WEBVTT
00:00:00.080 --> 00:00:04.160
A big misconception that people have is that we're like purely a browser agent.
00:00:04.400 --> 00:00:16.960
And actually, the way that DEX works is we take in context from the browser and then we do actions through the form of tool calling, like not through MCP for us, but we make custom tools for the power apps that we care about.
00:00:17.120 --> 00:00:21.839
And that's been a hundred times more efficient and accurate for our users.
00:00:22.160 --> 00:00:25.359
Most podcasts talk to founders after the story's already written.
00:00:25.519 --> 00:00:29.120
When the latest trade is announced, the product has shipped and the narrative is clean.
00:00:29.440 --> 00:00:32.880
We wanted to find out why they decided to take the founder leap in the first place.
00:00:33.039 --> 00:00:36.880
What idea captured their imagination and how they handle the pressure to build?
00:00:37.039 --> 00:00:38.079
Welcome to First Commit.
00:00:38.240 --> 00:00:39.200
I'm Laura Hamilton.
00:00:39.439 --> 00:00:42.320
I'm a DI loan and we're investors at Notable Capital.
00:00:42.479 --> 00:00:45.679
We spend a lot of time with technical founders who are on the cutting edge.
00:00:45.840 --> 00:00:49.679
If you're thinking about what it's like to start a company in the AI era, this one's for you.
00:00:49.920 --> 00:00:50.960
This is First Commit.
00:00:51.200 --> 00:00:52.399
Let's get into it.
00:00:53.920 --> 00:00:58.880
Today on First Commit, we have Regina Lynn, the CEO and co-founder of Dex.
00:00:59.280 --> 00:01:04.239
Dex is building browser agents that are trying to be the knowledge workers for everyone.
00:01:04.400 --> 00:01:05.599
Welcome, Regina.
00:01:05.920 --> 00:01:07.439
Thank you so much, Laura.
00:01:07.760 --> 00:01:14.159
So you went to undergrad at Harvard, and now you're building a browser agent from scratch where you started off at YC.
00:01:14.480 --> 00:01:19.280
Can you just touch on the moment that you realized the problem you're solving was real and not just a concept?
00:01:19.599 --> 00:01:22.159
Yeah, this is lots of pack already.
00:01:22.480 --> 00:01:28.560
So maybe to give you guys some context, I can start with when we first originally started working on the problem.
00:01:28.719 --> 00:01:34.319
I think for both my co-founder Kevin and I, it was very much a personal problem that we had.
00:01:34.480 --> 00:01:39.200
We were juggling a lot of work, undergrad, masters, building stuff on the side.
00:01:39.359 --> 00:01:42.079
And both of us were never like the super organized type.
00:01:42.159 --> 00:01:52.400
So like my whole life, I never previously never used Google Calendar, never had a notes app, tried Notion many times, and the nested pages just didn't do it for me.
00:01:52.640 --> 00:01:55.840
So, and and maybe even going a little bit before that.
00:01:56.000 --> 00:01:58.319
I grew up playing piano professionally.
00:01:58.480 --> 00:02:01.040
So did eight hours of practice a day.
00:02:01.359 --> 00:02:02.000
When did you start?
00:02:02.239 --> 00:02:02.959
When I was five.
00:02:03.599 --> 00:02:03.760
Yeah.
00:02:03.920 --> 00:02:10.560
So my whole life, I started practicing five hours when I was that, and then in high school, it became eight to ten hours.
00:02:10.719 --> 00:02:12.879
I was also doing public high school.
00:02:12.960 --> 00:02:17.759
So that meant I started my homework at midnight and had maybe two to three hours before it lights out.
00:02:18.000 --> 00:02:21.759
So because of that, I couldn't do digital chores with like organizing whatever.
00:02:21.840 --> 00:02:23.199
Like I just need to get the work done.
00:02:23.360 --> 00:02:25.360
And also everything was always in my head.
00:02:25.520 --> 00:02:29.520
So constantly tracking everything I had to do, juggling between those different things.
00:02:29.759 --> 00:02:36.560
And I think it was around our junior year in college when Kevin and I started working on a little side project.
00:02:36.800 --> 00:02:39.840
People still called everything GPT rappers at the time.
00:02:40.080 --> 00:02:42.400
And I think we were working really late night.
00:02:42.560 --> 00:02:48.800
So we wanted to have GPT 4.0 order DoorDash for us late at night when we forgot to eat dinner.
00:02:48.960 --> 00:02:51.599
And that was the very first like use case for ourselves.
00:02:51.680 --> 00:02:52.960
It was to do that.
00:02:53.120 --> 00:02:53.840
And it worked.
00:02:53.919 --> 00:03:02.319
So then after that, we had it do a variety of other things, like organizing our school work into calendar, then eventually applied it to like an invoicing use case.
00:03:02.479 --> 00:03:07.199
And I think that was the moment when we realized this was maybe like a useful thing for other people too.
00:03:07.439 --> 00:03:10.960
What happened was we snuck into this AI conference in Boston.
00:03:11.280 --> 00:03:12.000
Which one?
00:03:12.240 --> 00:03:13.680
I don't even remember the name.
00:03:13.759 --> 00:03:17.360
And we like found any random event that was happening outside of campus.
00:03:17.599 --> 00:03:24.479
And we had to print out fake lanyards and or name cards and buy lanyards to sneak into the conference too.
00:03:24.639 --> 00:03:32.639
But we had just so happened to bump into someone who he managed, I think, hundreds of like UI Path bots as his job.
00:03:32.800 --> 00:03:38.560
And he saw our demo just casually in the hallway of the of the vet and he was like super surprised.
00:03:38.639 --> 00:03:41.439
And he was telling us like this can replace UI Path.
00:03:41.520 --> 00:03:43.360
Like I have to manage all these computers.
00:03:43.520 --> 00:03:45.280
It always breaks, it doesn't do anything.
00:03:45.439 --> 00:03:51.120
And you guys have this piece of tech that can like just download the PDF and put it into QuickBooks.
00:03:51.439 --> 00:03:54.639
And that was what made us realize, like, hey, maybe this could be useful.
00:03:55.039 --> 00:03:56.719
So you were at Harvard at the time.
00:03:56.800 --> 00:03:58.879
And so did you end up finishing or did you drop out?
00:03:59.199 --> 00:04:01.120
I ended up leaving and dropping out.
00:04:01.199 --> 00:04:01.520
Yeah.
00:04:01.919 --> 00:04:09.280
Did you see any parallels from your years of piano playing to building or coding a new company?
00:04:09.520 --> 00:04:10.400
Yeah, yeah, for sure.
00:04:10.479 --> 00:04:18.240
I think like I was looking back and reflecting, and all those hours I spent really focusing on doing one thing super, super well.
00:04:18.399 --> 00:04:20.319
That definitely like taught me a lot.
00:04:20.399 --> 00:04:32.639
And I think it also led to me pivoting into math as my as my major in in college, because that's concentration where you really have to sit and think about a problem for a really long time.
00:04:32.800 --> 00:04:35.519
Um, and then after that, with building, it's the same thing.
00:04:35.600 --> 00:04:39.199
It's like you have to really commit and be really passionate about it.
00:04:39.360 --> 00:04:46.319
Yeah, like no kid wants to give up all those things repeatedly for something unless they're really, really passionate about that one thing, right?
00:04:46.560 --> 00:04:48.720
It's probably very rewarding for you, also.
00:04:48.800 --> 00:04:53.920
Like if you had all that energy to give to it, there had to be some feedback mechanism.
00:04:54.240 --> 00:05:01.680
Yeah, I I in terms of feedback, I think like the other thing that it taught me was like being really good at taking harsh feedback.
00:05:01.839 --> 00:05:11.360
So, like from piano teachers to public masterclass where you perform and you're getting roasted in front of the the crowd and they're there to take notes and watch you get critiqued.
00:05:11.519 --> 00:05:19.120
And then also competitions where it's very subjective and also you like your whole years of work gets put into one moment on stage.
00:05:19.279 --> 00:05:23.600
But yeah, I do think it's rewarding to have done that for sure.
00:05:23.920 --> 00:05:30.319
Do you think that you're able to show up stronger under pressure now because of your piano experience?
00:05:30.639 --> 00:05:38.240
Yeah, I was just thinking about this yesterday because I was talking, I was catching up with some founders about like high stress situations and running their day-to-day work.
00:05:39.120 --> 00:05:42.240
Yeah, I was thinking about like what is it that makes it stressful?
00:05:42.399 --> 00:05:51.120
And I was thinking for myself, I feel like there's been times where I was so stressed about to get on stage and perform in front of like 100 people.
00:05:51.279 --> 00:05:58.480
And I, for example, didn't have the piece fully memorized because I was doing school and I had maybe two weeks to memorize a 30-minute program.
00:05:58.720 --> 00:06:03.199
Like, what was I gonna do if I like forgot I had just memorized it the day before?
00:06:03.360 --> 00:06:11.600
And that kind of pressure of having to just get on there and fake it till you make it and just hope for the best and and make it out, like having performed something.
00:06:11.680 --> 00:06:18.800
Like I feel like nothing so far has been as daunting as actually doing that, even despite all the difficult decisions to make.
00:06:18.879 --> 00:06:23.920
Like, I know that I can face it head head on after after having done that before.
00:06:24.319 --> 00:06:26.079
There's something about performance, right?
00:06:26.160 --> 00:06:33.279
Like there's showing up in the moment and giving your best, even though you've kind of like prepared and it's never it's never what you imagined.
00:06:33.759 --> 00:06:37.839
Yeah, it never happens the way you expect it to or hope that it would be.
00:06:38.000 --> 00:06:38.240
Yeah.
00:06:38.560 --> 00:06:44.879
Do you remember a time, maybe music or later in life where you really had to like improvise?
00:06:45.360 --> 00:06:46.480
Yeah, yeah, for sure.
00:06:46.639 --> 00:06:49.600
Wait, so was it during music and also after?
00:06:49.839 --> 00:06:55.439
Yeah, one time I actually did forget the page, like the very last two minutes of a concerto.
00:06:55.680 --> 00:06:59.680
It was like the very last whole whole competition piece was like 30 minutes.
00:06:59.839 --> 00:07:02.959
At the very end, I had forgotten it and to improv the ending.
00:07:03.199 --> 00:07:06.879
And I just like made something up and like hit the last note and then walked off the stage.
00:07:06.959 --> 00:07:08.639
Luckily, it was contemporary.
00:07:08.879 --> 00:07:13.680
So when it didn't sound right, like who knows that it might be part of the part of the piece.
00:07:13.920 --> 00:07:14.720
Yeah, definitely.
00:07:14.800 --> 00:07:21.519
I think that has helped me a lot in terms of thinking out of the box when solving problems now.
00:07:21.680 --> 00:07:30.000
Like we operate in a very, I think, like dynamic space where every week something new comes out, you have to think about like first principles like, is what we're doing correct?
00:07:30.160 --> 00:07:31.519
Should we adapt to this right now?
00:07:31.600 --> 00:07:41.040
Or yeah, and and I think that flexibility of being able to handle any kind of situation depending on what's happening in the moment, that's something that's been helpful as well.
00:07:41.360 --> 00:07:44.160
How have you thought about doing that as a first-time founder?
00:07:44.560 --> 00:07:52.879
Like, how are you learning like when to iterate, what hypotheses to test, especially in an exciting but also competitive market that Dex is operating in?
00:07:53.040 --> 00:07:56.560
And it would probably be helpful to touch on Dex as well, just so we can ground folks.
00:07:56.800 --> 00:08:06.560
So Dex is a browser agent that lives inside of Chrome and helps people do manual work in their browser, and it also connects to a variety of integrations outside of the browser.
00:08:06.720 --> 00:08:11.920
So really helping people save time from all the manual admin digital chore work that they have to do.
00:08:12.160 --> 00:08:18.240
And I guess in terms of your question on learning, that's something I've kind of thought about a lot.
00:08:18.319 --> 00:08:25.519
It's like if you haven't seen something before, how do you figure out if you know everything you need to know to make the best decision right?
00:08:25.680 --> 00:08:39.440
And I think like having gone through building for the last year, a lot of times not nothing I can read or learn about from somebody else is as good as just having done it and heard the feedback directly from like your own actions.
00:08:39.679 --> 00:08:48.399
Because I think it's hard to compare situations or get advice because no one fully knows the exact situation and all the details of like your exact product.
00:08:48.559 --> 00:08:52.240
There might be some very tiny detail resulting in some results.
00:08:52.399 --> 00:09:00.080
So, like only you know the full picture, like have the full context on like what might be the actual reason for user feeling a certain way.
00:09:00.240 --> 00:09:11.120
And so I think like it's been a lot more fruitful for us to just do it instead of thinking too much, like trying to just put it out there and then get that feedback from from seeing what actually happens.
00:09:11.440 --> 00:09:13.200
So you describe Dex beautifully.
00:09:13.440 --> 00:09:18.559
Help us ground it a bit in like who is the target for it, information workers or where do they live today?
00:09:18.960 --> 00:09:24.720
So our ICP is someone who fully lives in the browser for like 100% of the day, pretty much.
00:09:24.960 --> 00:09:33.519
And when you think about like where in the industry or like the market those people are, we've tend to find them to be in go to market right now.
00:09:33.679 --> 00:09:42.879
And their entire work day is like in the browser inside of apps and also like a variety of tabs that don't have integrations, like LinkedIn, for example.
00:09:43.039 --> 00:09:51.679
And the other thing that is really interesting is I have been asking everybody I speak to like, do you have a personal to-do list anywhere?
00:09:51.840 --> 00:09:54.080
And for go-to-market people, they never do.
00:09:54.240 --> 00:09:56.879
They don't use the sauna, they don't use linear or anything like that.
00:09:57.120 --> 00:10:03.919
Sometimes their team tries to get them to, but I think the reason is their whole day evolves around other people and it's in the browser.
00:10:04.080 --> 00:10:12.399
So it's way too tedious to spend time putting down their to-dos when their to-dos are literally living like in their email, their calendar, their tabs.
00:10:12.639 --> 00:10:15.759
So that those are the people that we're serving right now.
00:10:16.000 --> 00:10:17.519
I actually spent a lot of time go to market.
00:10:17.759 --> 00:10:20.559
I used to run global operations for a public company.
00:10:20.879 --> 00:10:23.279
Spent a lot of time salespeople and go to market folks in general.
00:10:23.360 --> 00:10:24.720
And many times it's implicit.
00:10:24.799 --> 00:10:27.519
It's like their to-dos are the tabs they have open.
00:10:27.679 --> 00:10:28.000
Yeah.
00:10:28.240 --> 00:10:29.759
The tab being open is the reminder.
00:10:29.840 --> 00:10:32.080
Is the me with blog posts.
00:10:32.320 --> 00:10:35.279
If I open a blog post, they need to read it before the end of the day.
00:10:35.600 --> 00:10:35.759
Yeah.
00:10:35.919 --> 00:10:41.519
They I look at their browser and it's always like four windows maximum to the brim with tabs.
00:10:41.679 --> 00:10:48.879
I think like they try to share a screen with me, and then usually the browser crashes because they're waiting for it to load like every two clicks.
00:10:49.120 --> 00:10:50.320
So we see that a lot.
00:10:50.480 --> 00:10:51.840
And I was actually surprised.
00:10:52.000 --> 00:10:57.200
Like I we thought people had a lot of tabs, but I found people with like way more tabs than I expected.
00:10:57.519 --> 00:11:01.279
What has it been like acquiring early users or going to market?
00:11:01.519 --> 00:11:03.919
How are you finding folks to try ducks out?
00:11:04.320 --> 00:11:07.120
So far, it's been very like natural, word of mouth.
00:11:07.279 --> 00:11:10.799
We haven't invested too much into an actual go-to-market motion yet.
00:11:10.960 --> 00:11:22.240
And this is because we wanted to actually go for like really high retention or make sure we're adding a ton of value into the users who are naturally discovering us first before like growing more with a leaky bucket.
00:11:22.399 --> 00:11:26.399
And I think like as of the last few weeks, we're very happy now with our retention.
00:11:26.480 --> 00:11:28.639
So we'll be doing that more in the future.
00:11:28.879 --> 00:11:36.720
But I think like it's been great to use like LinkedIn and Instagram for us, just having our own team show our own use cases for the product.
00:11:36.879 --> 00:11:38.080
That's actually a big thing for us.
00:11:38.159 --> 00:11:44.159
Like, I'm still like three user of the product today, with like many thousand other people that use it quite a bit too.
00:11:44.399 --> 00:11:47.919
So, yeah, every second of my day, I'm I'm using the product.
00:11:48.240 --> 00:11:54.639
I wanted to go actually back to your point about discovering your your ICP.
00:11:54.799 --> 00:11:58.159
You start in a company, you have this idea, cool.
00:11:58.399 --> 00:12:04.080
One of the characteristics is like almost any human that's in the browser, which is almost any human, can use this.
00:12:04.240 --> 00:12:08.080
How did you end up narrowing it down to go-to-market folks or salespeople?
00:12:08.320 --> 00:12:15.679
Yeah, it's kind of a long story, but before my days building, I w dabbled a little bit in the investment banking world.
00:12:15.840 --> 00:12:22.720
And that was another shocking thing where I looked at Harvard, like, where did the smartest people I know, where were they going?
00:12:22.879 --> 00:12:28.000
And one of them happened to be a pianist, and he was at the time president of every finance club.
00:12:28.080 --> 00:12:29.600
So he was like, you should do this.
00:12:29.759 --> 00:12:31.919
And I was like, sure, this guy did what I did.
00:12:32.000 --> 00:12:33.519
Maybe I should go down that path.
00:12:33.679 --> 00:12:47.360
And during my summer internship and like the time I was kind of recruiting, talking to people in that space, so many very smart, ambitious young people that I knew in college, they left, they entered that industry and they came out just completely burnt out.
00:12:47.440 --> 00:12:53.200
Like they got sent to the hospital, even and every year there's there are people that die like in banking because they're sleeping.
00:12:53.440 --> 00:12:58.720
And then having done the job myself, one of the reasons I actually wanted to explore it was like, what are people doing?
00:12:58.879 --> 00:13:00.639
Like, what's taking that long?
00:13:00.720 --> 00:13:03.440
Like, what why is it like drinking from a fire hose?
00:13:03.600 --> 00:13:14.000
And I got a lot of those questions answered after having seen it, but also like at the same time, a lot of the work that the analysts and people are doing, it's really all automatable by agents today.
00:13:14.240 --> 00:13:23.279
And even I guess like in 2024, it was less possible, but still a lot of it, a lot of it could have been helped greatly tools and better technology.
00:13:23.440 --> 00:13:28.080
So I think I became very passionate right before building decks with that problem.
00:13:28.240 --> 00:13:32.080
I was like, people just should not be having to do this kind of work.
00:13:32.159 --> 00:13:42.080
And even outside of banking, just like regular admin work, there are many people who like every day their job is to do the same like monotonous task in their browser workspace.
00:13:42.240 --> 00:13:50.559
And I just think that if people were freed from doing that coordination work, from moving data between software and all of that, like the world would be a better place.
00:13:50.799 --> 00:13:59.360
So that was kind of what led me to explore where we would be able to kind of like benefit people the most.
00:13:59.600 --> 00:14:08.399
And I think like from there we went from finance to market research when we had applied to YC, which is tied to like one of the other questions.
00:14:08.639 --> 00:14:16.320
And eventually we realized that like we wanted to we started in the browser because that's where we can immediately get the most context.
00:14:16.559 --> 00:14:19.279
So you're able to get a ton of context from the browser.
00:14:19.600 --> 00:14:24.080
What are the capabilities today that you can achieve with that context versus what you can't?
00:14:24.399 --> 00:14:30.879
I think like when most people start out, especially students, they think that being on the desktop is like the best move.
00:14:30.960 --> 00:14:32.639
And this is what we originally did too.
00:14:32.799 --> 00:14:35.519
And very quickly you realize you have a couple paths forward.
00:14:35.600 --> 00:14:40.799
Like either you make a browser extension back before Web MCP came out, and to get more browser.
00:14:41.120 --> 00:14:42.720
You build a whole browser like some have done.
00:14:42.960 --> 00:14:43.279
Yeah.
00:14:43.519 --> 00:14:48.080
Or you have to spend a lot of time working with AppleScript or the Microsoft ecosystem.
00:14:48.320 --> 00:14:50.000
We actually like dabbled in all of those.
00:14:50.080 --> 00:14:53.679
And then we ended up picking the browser because the context is most available.
00:14:53.840 --> 00:14:57.200
There's a lot of tooling provided in the native Chrome ecosystem.
00:14:57.440 --> 00:15:05.600
And beyond just depending on vision, you can also utilize the DOM to have more structured clues into what the user is doing.
00:15:05.759 --> 00:15:07.360
So that was why we picked the browser.
00:15:07.519 --> 00:15:11.759
And I think like now back to the question of like how did we find our current ICP?
00:15:11.919 --> 00:15:17.519
It's really thinking about who depends on the browser as their command center when they work every day.
00:15:17.759 --> 00:15:26.799
And the kind of magic moment early on for our product was just not having to contact switch between their tabs or even just like the organized tabs feature.
00:15:26.960 --> 00:15:32.080
We had spent a lot of time building all these agentic features and like actions that people could do.
00:15:32.240 --> 00:15:41.200
But then at the end of the day, so many people they were like, oh, I just didn't have to switch to this other tab to grab a piece of information, or like, oh, I really like that it cleaned up my tabs for me.
00:15:41.360 --> 00:15:55.600
And that was when we started realizing, like, hey, maybe the value add here is that context switching piece, not like us trying to tell everyone like this can do actions for you, which I think until more recently, people didn't really understand how AI could even do actions for you.
00:15:55.919 --> 00:15:56.480
It's really interesting.
00:15:56.559 --> 00:16:03.039
I wonder if you guys have data on how many times on average people context switch between tasks.
00:16:03.440 --> 00:16:11.840
Yeah, we had this interesting idea to do like a browser wrapped because you can see how productive you were every day, which I think for knowledge workers did never really existed.
00:16:11.919 --> 00:16:18.559
But for engineers, you can check your GitHub history, you can even show people your GitHub, like I did this much engineering work.
00:16:18.720 --> 00:16:26.720
But for with what we've built out, it's actually possible to kind of look back at a diary of your work daily and see like I've been rotating through tabs.
00:16:26.879 --> 00:16:32.240
And what we discovered was like the browser kind of mirrors the mental model of the user.
00:16:32.320 --> 00:16:42.559
And you can tell by matching what their job and their work and their day-to-day looks like to their natural like organization style or mental model around how they organize their day-to-day.
00:16:42.720 --> 00:16:49.200
If you mix and match these things, there's like a finite amount of buckets of types of ways that people kind of like to do work.
00:16:49.360 --> 00:16:53.600
And it's interesting because then you can see like some people rotate through their tabs very quickly.
00:16:53.679 --> 00:16:57.679
Like they linear go through them and they will exit out the moment they stop meeting them.
00:16:57.840 --> 00:17:00.000
And they like cycle through so many tabs every day.
00:17:00.080 --> 00:17:03.279
And then other people leave them all open and they're very attached to the tabs.
00:17:03.519 --> 00:17:10.160
I'm kind of like that, where the the tabs are like reminders for them and they always need to kind of jump to another task, come back.
00:17:10.240 --> 00:17:12.079
So all these threads are just like open there.
00:17:12.240 --> 00:17:19.440
And then you have people who are like very organized by windows, so they'll open up five different windows, and every workflow is in its own little window.
00:17:19.519 --> 00:17:25.039
And I think like Arc users are very much like that because they can organize their spaces easier.
00:17:25.359 --> 00:17:29.759
How do current users evaluate success metrics of the product?
00:17:30.079 --> 00:17:37.440
A successful day of using Dex or like first moment of using Dex is if it was able to take your mind off something.
00:17:37.599 --> 00:17:47.440
So, an example of that, like let's say something as simple as sending an email follow-up or scheduling, you have to pay attention and do that task, even if it's very simple, like responding.
00:17:47.519 --> 00:17:56.160
If your day is already filled back to back with stuff, like you that that extra effort of writing out that response and thinking about it, that is effort for people.
00:17:56.319 --> 00:18:01.759
Um, and that's why they like will switch to GPT, paste it in, and then like have GPT give you a response sometimes.
00:18:01.920 --> 00:18:04.160
They're just kind of tired of that mental effort.
00:18:04.319 --> 00:18:13.359
With Dex, we're trying to increase the difficulty and like the amount of time saved through taking their mind completely off of a task.
00:18:13.440 --> 00:18:17.119
So I can just tell Dex to schedule for me and don't have to think about it anymore.
00:18:17.279 --> 00:18:28.000
Um, they do have to come back to eventually approve some like very mission critical actions, but that whole in-between, like we want them to not have to think about it and just let the agent handle it.
00:18:28.319 --> 00:18:33.359
It's a meta pattern I love to call it offloading your bio tokens into digital tokens.
00:18:33.680 --> 00:18:35.599
Think about how many you're burning a day.
00:18:35.759 --> 00:18:36.000
Yeah.
00:18:36.240 --> 00:18:37.759
These reminders to close your day.
00:18:37.839 --> 00:18:40.640
And if yeah, to your point, easily automatable.
00:18:41.039 --> 00:18:45.279
I feel like 2026 is the year of these browser and computer use agents.
00:18:45.519 --> 00:18:51.200
I think we're seeing the model foundation labs as well as other startups just move deeper into this area.
00:18:51.359 --> 00:18:55.519
How do you think about the landscape and how to quote unquote stay ahead?
00:18:55.839 --> 00:18:57.599
I think a couple of things have changed.
00:18:57.759 --> 00:19:06.319
So, like one, when we started in 2024, fall, computer use had just come out and that became a term like in the very early days.
00:19:06.559 --> 00:19:12.079
And it was very bad, like it was just very slow and it wasn't very accurate.
00:19:12.240 --> 00:19:19.519
And I think that was the case for up until like literally this year in the last few months, it started getting faster and more accurate.
00:19:19.680 --> 00:19:25.599
And I think benchmarks are now around like 80%, and you can further improve that on specific data, right?
00:19:25.759 --> 00:19:37.119
So before everyone was thinking like you can only use this tech really for like full end-to-end automation and like a very traditional, like let's say logistics and go for the UI path route.
00:19:37.200 --> 00:19:41.920
And a lot of people who had been building browser agents actually wanted to use it for consumer.
00:19:42.000 --> 00:19:46.000
And I think also a lot of them were coming from a very engineering heavy perspective.
00:19:46.240 --> 00:19:54.319
Once they actually deployed these, they realized like knowledge workers don't really want to use this in the way that I thought they would want to use this for their work.
00:19:54.559 --> 00:19:56.880
So I think like now it's just that's one piece.
00:19:56.960 --> 00:19:59.759
Like it's a lot more possible, it's a lot faster, it's a lot more accurate.
00:19:59.920 --> 00:20:08.319
And then the other thing is I think there was a very big like mental model shift around going away from this full automation mindset to like co-working is okay.
00:20:08.400 --> 00:20:17.680
And like the whole co-working paradigm is like fine, and having human in the loop is necessary right now and actually improves like the agent's ability to learn and get better too.
00:20:17.839 --> 00:20:26.720
So I think that happened along with the launches of like cowork and the whole claude bot open cloud craze that went very viral.
00:20:26.880 --> 00:20:41.680
And I think those things really helped the people outside of maybe like the SF bubble or people who are less technical to realize like, actually, AI is not just for research and not just for summarizing, but it can do actual actions for you in your like real-world day-to-day tasks.
00:20:41.759 --> 00:20:51.519
And that that's been huge because you can see now everyone's looking for solutions to their problems and they're able to be creative about like what they want AI to actually do for them.
00:20:51.680 --> 00:20:59.119
So, yeah, I think those things combined made browser agents a lot more viable this year and a very exciting space to be in.
00:20:59.440 --> 00:21:04.000
It feels like you know, if we expand on it a bit to your point, like what people are actually trying to solve with this.
00:21:04.720 --> 00:21:11.039
Like almost every day, some playwright CLI came a few weeks ago, and now Alibaba released their page agent.
00:21:11.279 --> 00:21:13.680
And today I saw a couple other things that I can't even recall.
00:21:13.920 --> 00:21:18.160
The ecosystem is exploding, and you guys have been very early on the browser.
00:21:18.480 --> 00:21:23.200
Do you think are you married to the idea of the browser being the right space?
00:21:23.440 --> 00:21:28.079
Or it's more about the purpose, the thing you're trying to resolve and you will adapt to what's happening in the market.
00:21:28.319 --> 00:21:40.799
I think for us, we started on the computer and then we chose the browser because we were super obsessed with like how can we become the player who has captured the most context from the user and truly understand their whole workspace end to end.
00:21:41.039 --> 00:21:42.960
That was the reason why we picked the browser back then.
00:21:43.119 --> 00:21:47.519
There was a lot less standard tooling to your point with like playwright CLI coming out and all that.
00:21:47.599 --> 00:21:56.000
It's showing that like it's becoming more democratized and very standard to be able to access tooling in the browser and that agents need access to the browser.
00:21:56.160 --> 00:21:59.759
So a lot of helpful tooling has come out, but we are definitely not making.
00:22:00.160 --> 00:22:02.079
Married and we're not like browser only.
00:22:02.240 --> 00:22:07.279
Even today, a big misconception that people have is that we're like purely a browser agent.
00:22:07.440 --> 00:22:20.000
And actually, the way that DEX works is we take in context from the browser and then we do actions through the form of tool calling, like not through MCP for us, but we make custom tools for the power apps that we care about.
00:22:20.160 --> 00:22:24.799
And that's been a hundred times more efficient and accurate for our users.
00:22:25.039 --> 00:22:28.400
And we definitely also mixed in browser tooling with that.
00:22:28.480 --> 00:22:37.359
And that was something that we didn't see labs or like people doing that often for quite some time in the last year until more recently is like combining the two together.
00:22:37.519 --> 00:22:47.519
So that means the agent knows to call a browser tool and click on your screen when that's needed, and then it will call like directly into Gmail, for example, when that's more efficient.
00:22:47.839 --> 00:22:52.880
Have you guys seen like Claud and Chrome extension and like Atlas and all those?
00:22:53.039 --> 00:22:53.359
Yeah.
00:22:53.519 --> 00:23:02.880
If you try asking it to fill out a spreadsheet for you, it will take seven minutes plus to try to fill out one row, and then it ends up putting it into one cell.
00:23:03.039 --> 00:23:13.519
And that's just like the most obvious example for people when like you ask Dex to do it, it will be able to fill like hundreds of rows just in a few minutes accurately.
00:23:13.680 --> 00:23:16.640
And I think like the tool calling stuff is very helpful there.
00:23:17.039 --> 00:23:20.880
My agents have been playing a bit with the playwright CLI and the Alibaba agent of payages.
00:23:23.920 --> 00:23:27.519
What do you think are the bigger like constraints in the space right now?
00:23:27.759 --> 00:23:29.920
2026 compared to 2027.
00:23:30.079 --> 00:23:30.880
Where will we go?
00:23:31.200 --> 00:23:37.440
I think right now a lot of attention is on like how to create, gather, and utilize like team context.
00:23:37.519 --> 00:23:48.319
So it's no longer only about like your personal agent, but it's also like how can you share these agents to your team and have this collective data that's been extracted and utilized to make your agents better?
00:23:48.480 --> 00:23:52.799
And it's interesting because I think a lot of people want to build it from scratch right now.
00:23:52.880 --> 00:23:55.359
Like obviously, Cloud Codes made things a lot easier.
00:23:55.519 --> 00:24:05.200
And I think it's like not sustainable long term because it's very like one-off, and you are quite literally building a whole platform end-to-end if you want to make it something your whole team can use.
00:24:05.279 --> 00:24:06.880
Like there's a lot of pieces to learn.
00:24:06.960 --> 00:24:11.200
And when I talk to especially like go-to-market folks, they're already working till like 12.
00:24:11.599 --> 00:24:14.640
Like they have these back-to-back calls, and then they have to learn how to code.
00:24:14.720 --> 00:24:20.160
They have to learn how context works, like how agents work, and then they're trying to vibe code a solution to build agents.
00:24:20.319 --> 00:24:22.319
And it's just a lot, a lot to do.
00:24:22.480 --> 00:24:33.680
And so I think like looking at the year ahead, a lot of the interesting part is around like one, how do you actually get structure the context, retrieve it, and then allow agents to use it?
00:24:33.839 --> 00:24:39.839
And then another big thing we're very focused on is like what's the correct interface or like way to interact with the user.
00:24:39.920 --> 00:24:46.480
I think a lot of it's actually figuring out the right scaffolding around the tech because the tech works now, it's only going to get better.
00:24:46.640 --> 00:24:51.759
But allowing people to actually utilize it in the mental model that they that makes sense to them.
00:24:51.920 --> 00:24:56.000
I think that piece is like under under figured out at the moment.
00:24:56.319 --> 00:25:01.839
So like kind of different people want to use different agents in different places, right?
00:25:01.920 --> 00:25:04.240
And I think you should be able to use it from anywhere.
00:25:04.400 --> 00:25:08.960
So some people might work from the browser, they'll use it from the browser CLI phone.
00:25:09.119 --> 00:25:13.839
Depending on where their day is, like that's where they should be able to use the technology.
00:25:14.160 --> 00:25:17.839
And then there's also a lot of questions with like how do you make sure it's reliable?
00:25:18.000 --> 00:25:22.400
How do you build in observability and all of that is very important to make it truly useful?
00:25:22.960 --> 00:25:28.960
Yeah, as you're building into the into businesses, into the enterprise, obviously trust and governance and yeah.
00:25:29.759 --> 00:25:31.759
How do you guys think about like parallelization?
00:25:32.480 --> 00:25:35.039
Which is notoriously complicated.
00:25:35.279 --> 00:25:38.480
It's hard to drive the browser multiple directions at the same time.
00:25:39.039 --> 00:25:42.559
So we have parallel agents running, calling the tools.
00:25:42.799 --> 00:25:51.680
So for example, if I wanted to reach out with custom personalized emails to this whole spreadsheet of people, we'll send off different agents to go do that task.
00:25:51.839 --> 00:25:55.440
And they just the user just sees these pop up like in their browser sidebar.
00:25:55.759 --> 00:26:02.880
So for difficult tasks like that that are in bulk, we're not sending off browser agents because that will kill their Chrome immediately.
00:26:03.119 --> 00:26:06.240
On the other hand, we play around with the tab groups a lot.
00:26:06.319 --> 00:26:15.359
So we've been trying to see like if we can figure out this tab group or like these scattered tabs are one work stream and one thread, and these other ones are a different one.
00:26:15.440 --> 00:26:27.279
And if we can first of all like tab group things correctly, not by topics of the tab, but by their actual workflow, then can we also have a different sidebar come up to tackle that task in that tab group?
00:26:27.440 --> 00:26:29.039
So this is something we're exploring with.
00:26:29.200 --> 00:26:42.160
As of right now, the way it works is it's one universal sidebar for like all the then people have actually liked that more than some of the AI browsers, where it opens up a new sidebar per tab because then you're still contact switching a lot.
00:26:42.559 --> 00:26:43.279
That makes sense.
00:26:43.519 --> 00:26:45.119
How many people are on the team today?
00:26:45.359 --> 00:26:46.480
It's still very small.
00:26:46.640 --> 00:26:48.480
So we have four people full-time.
00:26:48.640 --> 00:26:51.279
It's my co-ventor and I, and then two engineers.
00:26:51.599 --> 00:26:56.400
How have you thought about building an AI-first product with in an AI-first world?
00:26:56.559 --> 00:26:58.319
Are you able to keep a leaner team?
00:26:58.720 --> 00:27:05.119
Our internal goal right now is actually to run all of our operations on DEX by the end of April.
00:27:05.279 --> 00:27:14.319
So currently we have two growth interns, and their job before they leave is to like automate their own work as much as possible.
00:27:14.480 --> 00:27:20.640
And we've been feeding all of our context to our context engine, which DEX also connects to and runs off of.
00:27:20.799 --> 00:27:23.519
It's been like amazing how much it can do for me.
00:27:23.680 --> 00:27:30.799
So, like to give you a tangible example, I was evaluating CRMs in the last little bit because I'm now talking to too many different customers.
00:27:30.960 --> 00:27:34.400
They're telling me different feature requests and they all have different blockers.
00:27:34.640 --> 00:27:45.839
And I couldn't find really a CRM that's far enough where it can actually gather all of my data from everywhere and then surface the right results with like AI to analyze and consolidate that.
00:27:46.079 --> 00:27:57.519
And then after we had connected our contract system to all these integrations that we have, so it's live updating with every Slack message, every email I'm getting, every call, calls that I'm not in as well.
00:27:57.680 --> 00:27:59.519
And all of that's feeding in live.
00:27:59.599 --> 00:28:05.279
And at any moment, Dex has access to that and knows and has this complete understanding of what's going on.
00:28:05.519 --> 00:28:11.759
So basically, what I asked it to do was just like generate me a spreadsheet with all the fields that I care about.
00:28:11.839 --> 00:28:16.240
Like I want to know what's blocking every customer right now that I spoke to in the last week.
00:28:16.400 --> 00:28:21.839
I also want to know what feature requests have been most requested and what should be prioritized in our product roadmap.
00:28:21.920 --> 00:28:24.559
And it was able to do that in I could do that like one minute.
00:28:24.640 --> 00:28:27.519
And it was more comprehensive than any CRM that I tested.
00:28:27.680 --> 00:28:30.799
And I can just whip up an analysis like that at any given moment.
00:28:30.960 --> 00:28:32.559
So that's been super powerful.
00:28:32.720 --> 00:28:34.960
And that's not even an automation use case.
00:28:35.039 --> 00:28:40.799
Like I can have this run on a trigger, I can do a bunch of things to tie like all the agents in the background together.
00:28:40.880 --> 00:28:42.960
So yeah, it's definitely very powerful.
00:28:43.200 --> 00:28:46.079
And we are planning to run our company off of that.
00:28:46.319 --> 00:28:47.839
That sounds a lot like a memory problem.
00:28:48.000 --> 00:28:52.559
That feeds into your context engine, or are you guys using something from the community for that?
00:28:52.960 --> 00:28:54.720
We build everything from scratch.
00:28:54.880 --> 00:28:58.319
So, how the way that we handle context and memory, there's two pieces.
00:28:58.559 --> 00:29:13.920
One is kind of the source of truth that we're extracting from integrations, and that just comes from having a lot of connectors, extracting the data out, and building relationships and understanding what is this company's source of truth or like their reality, essentially.
00:29:14.079 --> 00:29:20.960
And then the other piece comes from observing in the browser what the daily processes and decisions are being made coming from each individual.
00:29:21.119 --> 00:29:27.519
And we combine these two together to actually like enrich the knowledge of what Dex can understand and know.
00:29:27.680 --> 00:29:34.400
And the two pieces kind of give you different pros and cons, but it builds a much better understanding of everything.
00:29:34.799 --> 00:29:40.799
How far do you think you'll be able to scale as a company with call it four, five, six people?
00:29:41.599 --> 00:29:42.960
I think you can do a lot.
00:29:43.119 --> 00:29:48.319
Like what I realized is that software is easy to make now.
00:29:48.480 --> 00:29:57.759
Like the like the CRM piece or like a dashboard, or if you have full access to your data and you can utilize it with agents, it can do so much for you.
00:29:57.839 --> 00:29:59.599
It can just, you can just vibe code a CR.
00:29:59.839 --> 00:30:07.519
If you wanted one, you can vibe code a dashboard or an automation or an agent, and all of it can kind of do work for you when you're resting and sleeping.
00:30:07.839 --> 00:30:11.039
So I think like that's the direction I see things going.
00:30:11.119 --> 00:30:13.680
It's almost like lovable, but for agents.
00:30:13.839 --> 00:30:20.160
And what needs to happen though is like you need to be able to extract and understand all the data that you have.
00:30:20.240 --> 00:30:36.960
And then there's a missing piece of data, which is what we're trying to figure out how to collect more of, which is kind of like how are you people making decisions and what actually results in a certain outcome that gets logged somewhere versus just looking at the outcome and not knowing how that got there.
00:30:37.359 --> 00:30:40.720
Building on that, you're a very lean team, you're achieving a lot, you're moving really fast.
00:30:40.960 --> 00:30:50.319
Is there some magic sauce in how you organize your engineering best practices or things that you've developed along the way to help you to do that?
00:30:50.640 --> 00:30:50.799
Yeah.
00:30:50.960 --> 00:30:57.519
So with building off of what I said earlier with the CRM that I can just create at any second, I have that update.
00:30:57.680 --> 00:31:01.759
Every time I have new conversations after a few days, I'll ask it to update that.
00:31:01.920 --> 00:31:08.240
And then with that, I immediately tell it to prioritize like the product roadmap items we have and add it all to our linear.
00:31:08.400 --> 00:31:13.200
So that piece, no one has to kind of manually add tickets into linear.
00:31:13.440 --> 00:31:21.200
Anytime anyone gets like a customer request, either through a call, an email, or in Slack, we also just ask Dex to go add it into Linear.
00:31:21.440 --> 00:31:23.920
So all of that work saves so much time for us.
00:31:24.079 --> 00:31:27.519
And you can also, that's like the day-to-day piece.
00:31:27.680 --> 00:31:32.160
You can also kind of take a step back and it helps me analyze like the direction that we're going in.
00:31:32.319 --> 00:31:40.160
So I use it a lot for brainstorming and looking at all the data at once, which is not something you can humanly do by yourself.
00:31:40.240 --> 00:31:43.440
Like that's I think one of the greatest use cases for AI.
00:31:43.680 --> 00:31:56.079
I can have it read through every single interaction that we've had in the last few months, or like look through logs and figure out, like on the broader scheme of things, what are we getting wrong in our assumptions of our users?
00:31:56.240 --> 00:32:00.799
Like, is is what we think correct, or is there something else that's going on here?
00:32:00.960 --> 00:32:05.440
So constantly like analyzing and reanalyzing and figuring it that out.
00:32:05.599 --> 00:32:08.640
And I think that's something I would like very much recommend as well.
00:32:08.960 --> 00:32:16.319
Maybe we could do one last subject around where Dex is today, like what's keeping you up at night?
00:32:16.880 --> 00:32:27.119
I think one big thing that's a little bit surprising, and actually it's not that surprising once you think about it, is that one of the blockers at the end of the day is human creativity.
00:32:27.279 --> 00:32:36.640
Like people are not very good at systems thinking or managing others, and giving them agents is essentially telling them you have this team that you can go tell them to do whatever.
00:32:36.880 --> 00:32:41.759
And what we found is that people find it hard to come up with new use cases.
00:32:41.839 --> 00:32:52.480
So even if we give them templates or we give them suggestions, sometimes they're like, oh great, I can use decks for this one use case, and they'll always use it for that one use case, but they're not able to think of more use cases.
00:32:52.640 --> 00:32:54.559
So that was like a big learning for us.
00:32:54.880 --> 00:33:01.039
It's sometimes better to educate your user on the capability rather than a feature.
00:33:01.200 --> 00:33:04.640
And a lot of times people will take use cases as features.
00:33:04.720 --> 00:33:13.279
They're like, okay, I can do this use case, but that to them that's a feature rather than like, I understand the capabilities underneath and I can now apply it to other situations.
00:33:13.519 --> 00:33:26.079
So we've been working a lot on kind of like the education piece, and I think that's actually a plus when it comes to sales because many teams right now, they're also internally trying to figure out is the right move to adopt a product?
00:33:26.240 --> 00:33:33.759
Is it to build internal tooling, or should we actually invest in like a cloud code program to teach everyone how to use AI and how to build agents?
00:33:33.920 --> 00:33:43.599
I guess like education is a big piece of what we're doing, and we're like contributing to helping people actually understand how to use agents and like what's going on underneath the surface.
00:33:43.920 --> 00:33:46.319
That's such a beautiful, you've crystallized it very well.
00:33:46.480 --> 00:33:52.799
I find that AI is fundamentally very new and unlocked a lot of things, but also in many ways it's not new at all.
00:33:53.039 --> 00:33:56.559
To your point, the limiting factor seems to be thinking and levers.
00:33:56.799 --> 00:34:02.480
Humans who can understand the levers of a system have found ways historically how to achieve leverage, right?
00:34:02.559 --> 00:34:04.559
Hiring other people, building systems.
00:34:04.720 --> 00:34:07.599
Now it's kind of like AI is bringing leverage to anyone.
00:34:07.759 --> 00:34:07.920
Yeah.
00:34:08.079 --> 00:34:12.320
And if you don't have that thinking, you're just gonna phase zero interact with it back and forth.
00:34:12.400 --> 00:34:14.639
But if you think in a system, you can do so much.
00:34:14.960 --> 00:34:18.559
Yeah, and that's also why I brought up kind of like the lovable point, right?
00:34:18.639 --> 00:34:22.239
Because when websites first came out, it was difficult to make one.
00:34:22.320 --> 00:34:25.519
And everyone wanted to prove the fact they could build it from scratch.
00:34:25.599 --> 00:34:31.119
And then more and more tooling came out, and now we have like you can vibe code a full stack app or a website super easily.
00:34:31.199 --> 00:34:44.880
So I think the same thing will eventually happen with like building systems and agents, hopefully, where you don't have to think about as much going on behind the scenes and like the tooling that we're building and other people are building kind of take care of a lot of that process for you.
00:34:45.039 --> 00:35:00.000
So, like everything from context management to prompting correctly to routing to the right type of agent to use, like whether that's browser or like tool calling or all of that is kind of knowledge that it's important to have that knowledge to know how to optimize your workflow.
00:35:00.159 --> 00:35:08.960
And I realized that the reason why some users aren't able to use it as much as I can is because they don't have that knowledge and they wouldn't know that.
00:35:09.119 --> 00:35:10.159
Why would they know that?
00:35:10.320 --> 00:35:13.599
That's like something that will just improve gradually throughout the year.
00:35:13.920 --> 00:35:14.480
Meta, right?
00:35:14.639 --> 00:35:21.199
Like goes all the way back to that advice you got, that school contact that led you down the path to investment banking.
00:35:21.360 --> 00:35:25.199
And only through the pain you realize what you want to do and you understand the system.
00:35:25.280 --> 00:35:27.599
And now you built like a system of leverage all around that.
00:35:27.760 --> 00:35:28.159
Yeah.
00:35:28.400 --> 00:35:30.239
You have to go through the pain to learn.
00:35:30.480 --> 00:35:39.360
And that's actually also we didn't touch on this, but why we pivoted inside of our YC batch, we just realized that there was nothing wrong with the idea we were doing.
00:35:39.599 --> 00:35:40.159
What was it?
00:35:40.400 --> 00:35:42.719
We joined YC building customer intelligence.
00:35:42.800 --> 00:35:46.000
So this was around when we were really obsessing over context.
00:35:46.079 --> 00:36:00.320
And then we had this idea that what if you could have your product collect user touch points directly and use the interface to interact with users, collect those that data, and utilize that to adapt your website or adapt your product.
00:36:00.559 --> 00:36:05.599
And the reason why we moved out was purely kind of like a founder product fit reason.
00:36:05.760 --> 00:36:15.840
Like we wanted to build the original side project we were building back at school and not something because we thought like this is what people wanted or like it's B2B and all these other reasons.
00:36:16.000 --> 00:36:23.280
I think I actually had like a 30 to 50% response rate when I was messaging CMOs about the like customer market intelligence tool.
00:36:23.440 --> 00:36:35.199
But even that was not like as interesting to us than finding something that was kind of a mix between getting to build at the forefront of the tech, but also something that immediately people, ourselves and people around us like wanted to actually use.
00:36:35.360 --> 00:36:38.960
I think that's pretty rare of an idea to find and have that timing.
00:36:39.119 --> 00:36:42.000
So we couldn't really, we didn't want to lose that opportunity.
00:36:42.320 --> 00:36:42.719
For sure.
00:36:42.880 --> 00:36:46.480
I mean, it's really impressive that you pivoted and it's a long journey.
00:36:46.639 --> 00:36:51.280
So I I think it makes sense to do something that you're passionate about, especially with people that you're so close to.
00:36:51.440 --> 00:36:54.880
I bet you and your co-founder have such an authentic open relationship.
00:36:55.280 --> 00:37:04.000
I can't imagine not knowing, like having to work with someone where you don't know them for like very long, because then you never know what they're really like thinking.
00:37:04.079 --> 00:37:09.599
Whereas, like with my co-founder, we've been friends for so long, we've worked in many different situations for so long.
00:37:09.760 --> 00:37:15.599
Like, if he walks in the room after talking to a potential hire customer, I immediately get the vibe of what's going on.
00:37:15.760 --> 00:37:17.920
And that's like super nice to be able to have.
00:37:18.320 --> 00:37:21.280
Do you and your team you do you guys all live in the same house too?
00:37:21.599 --> 00:37:24.239
We used to all live together as well with our early employees.
00:37:25.840 --> 00:37:27.440
I think that's a very San Francisco thing.
00:37:27.519 --> 00:37:30.960
I spend my time between San Francisco and Tel Aviv that is not happening in Tel Aviv.
00:37:31.280 --> 00:37:33.599
Yeah, it's a little less frequent in New York, I would say.
00:37:34.320 --> 00:37:36.719
It's definitely an interesting experience.
00:37:36.960 --> 00:37:39.840
We all now moved out to be right across the street.
00:37:39.920 --> 00:37:41.840
So it's a 10-second walk for all of us.
00:37:42.000 --> 00:37:44.480
And I think it does make it evolving.
00:37:44.639 --> 00:37:44.800
Okay.
00:37:45.039 --> 00:37:46.719
We have our own doors now.
00:37:47.039 --> 00:37:49.840
Yeah, and we needed more meeting space in our office.
00:37:50.000 --> 00:37:54.480
So it's it's been nice to get like that sunlight when you walk to work in the morning.
00:37:54.800 --> 00:37:56.960
Sunlight seems to be important for humans.
00:37:57.199 --> 00:37:58.239
Yeah, for sure.
00:37:58.480 --> 00:38:00.480
Well, do you want to wrap with a quick fire round?
00:38:00.639 --> 00:38:01.199
Yeah, sure.
00:38:01.360 --> 00:38:01.599
Okay.
00:38:01.760 --> 00:38:05.039
What is one daily task you'll never do again in two years?
00:38:05.280 --> 00:38:08.960
Because it's two years, I'm gonna say something kind of extreme.
00:38:09.199 --> 00:38:14.559
I think by in in two years, I don't want to open a web app and use it like at all.
00:38:14.800 --> 00:38:18.159
I already like have gotten rid of a lot of that in my day-to-day.
00:38:18.320 --> 00:38:20.880
So I don't have to manually respond to emails.
00:38:20.960 --> 00:38:26.320
I don't have to schedule, like even opening Slack, updating linear, all of that's been like done through through Dex.
00:38:26.559 --> 00:38:31.119
And I think like in in two years, I wouldn't really need to like manually use the browser at all.
00:38:31.440 --> 00:38:35.119
What's the most underrated consumer AI use case right now?
00:38:35.440 --> 00:38:40.480
I think it'd be actually reminding you to do something or reminding you not to do something.
00:38:40.559 --> 00:38:43.920
So kind of like proactive, like keeping you focused.
00:38:44.079 --> 00:38:46.480
That's been something that's been really helpful and fun.
00:38:46.639 --> 00:38:48.159
And it's also a little bit gamified.
00:38:48.320 --> 00:38:58.880
Like if you have trouble with getting distracted, it can let you know, like, hey, you need to focus on whatever you're working on, or like the opposite, where maybe you forgot to put something into your system of record.
00:38:58.960 --> 00:39:05.760
Like you wouldn't really know that until like now you have this agent that can tell you you forgot to follow up or there's a task that is about to be due.
00:39:05.920 --> 00:39:08.480
These have been really like fun, interesting use cases.
00:39:08.800 --> 00:39:11.920
What's a non-obvious constraint in building agents?
00:39:12.239 --> 00:39:14.719
Definitely the interface piece of it.
00:39:14.800 --> 00:39:23.199
I think this is under talked about because it's always about like the technicalities of it, but actually like the interface itself, I think is really important.
00:39:23.360 --> 00:39:30.159
And it's because agents fundamentally operate around like your mental model and that includes like the interface you're using it out of.
00:39:30.239 --> 00:39:36.880
And when the CLI or like the IDE is not your natural interface, what should it look like?
00:39:36.960 --> 00:39:40.960
It can take many different form factors, you can interact with it in many different ways.
00:39:41.199 --> 00:39:44.320
Last one, what are you personally too stubborn to let AI do?
00:39:44.639 --> 00:39:50.639
Writing, like everything I write, I still will always make like a first draft of it or speak.
00:39:50.800 --> 00:40:02.880
I I will actually write out like my raw thoughts, even if it's not organized or cohesive, and then I will use AI to come up with more ideas to phrase things, but I don't want it to come up with the first copy for anything.
00:40:03.199 --> 00:40:13.039
The other thing is, oh, I another one that I'm huge, I'm very like, I think like AI interviews or anything where it replaces like you actually talking to someone face to face.
00:40:13.199 --> 00:40:15.679
There's just so much nuance that would get lost.
00:40:15.760 --> 00:40:18.719
Like, I could not imagine ever letting that go.
00:40:18.960 --> 00:40:19.679
I agree.
00:40:19.840 --> 00:40:20.800
Thanks, Regina.
00:40:21.280 --> 00:40:22.239
Thank you guys so much.
00:40:22.400 --> 00:40:23.039
Yeah.
00:00:00.080 --> 00:00:04.160
A big misconception that people have is that we're like purely a browser agent.
00:00:04.400 --> 00:00:16.960
And actually, the way that DEX works is we take in context from the browser and then we do actions through the form of tool calling, like not through MCP for us, but we make custom tools for the power apps that we care about.
00:00:17.120 --> 00:00:21.839
And that's been a hundred times more efficient and accurate for our users.
00:00:22.160 --> 00:00:25.359
Most podcasts talk to founders after the story's already written.
00:00:25.519 --> 00:00:29.120
When the latest trade is announced, the product has shipped and the narrative is clean.
00:00:29.440 --> 00:00:32.880
We wanted to find out why they decided to take the founder leap in the first place.
00:00:33.039 --> 00:00:36.880
What idea captured their imagination and how they handle the pressure to build?
00:00:37.039 --> 00:00:38.079
Welcome to First Commit.
00:00:38.240 --> 00:00:39.200
I'm Laura Hamilton.
00:00:39.439 --> 00:00:42.320
I'm a DI loan and we're investors at Notable Capital.
00:00:42.479 --> 00:00:45.679
We spend a lot of time with technical founders who are on the cutting edge.
00:00:45.840 --> 00:00:49.679
If you're thinking about what it's like to start a company in the AI era, this one's for you.
00:00:49.920 --> 00:00:50.960
This is First Commit.
00:00:51.200 --> 00:00:52.399
Let's get into it.
00:00:53.920 --> 00:00:58.880
Today on First Commit, we have Regina Lynn, the CEO and co-founder of Dex.
00:00:59.280 --> 00:01:04.239
Dex is building browser agents that are trying to be the knowledge workers for everyone.
00:01:04.400 --> 00:01:05.599
Welcome, Regina.
00:01:05.920 --> 00:01:07.439
Thank you so much, Laura.
00:01:07.760 --> 00:01:14.159
So you went to undergrad at Harvard, and now you're building a browser agent from scratch where you started off at YC.
00:01:14.480 --> 00:01:19.280
Can you just touch on the moment that you realized the problem you're solving was real and not just a concept?
00:01:19.599 --> 00:01:22.159
Yeah, this is lots of pack already.
00:01:22.480 --> 00:01:28.560
So maybe to give you guys some context, I can start with when we first originally started working on the problem.
00:01:28.719 --> 00:01:34.319
I think for both my co-founder Kevin and I, it was very much a personal problem that we had.
00:01:34.480 --> 00:01:39.200
We were juggling a lot of work, undergrad, masters, building stuff on the side.
00:01:39.359 --> 00:01:42.079
And both of us were never like the super organized type.
00:01:42.159 --> 00:01:52.400
So like my whole life, I never previously never used Google Calendar, never had a notes app, tried Notion many times, and the nested pages just didn't do it for me.
00:01:52.640 --> 00:01:55.840
So, and and maybe even going a little bit before that.
00:01:56.000 --> 00:01:58.319
I grew up playing piano professionally.
00:01:58.480 --> 00:02:01.040
So did eight hours of practice a day.
00:02:01.359 --> 00:02:02.000
When did you start?
00:02:02.239 --> 00:02:02.959
When I was five.
00:02:03.599 --> 00:02:03.760
Yeah.
00:02:03.920 --> 00:02:10.560
So my whole life, I started practicing five hours when I was that, and then in high school, it became eight to ten hours.
00:02:10.719 --> 00:02:12.879
I was also doing public high school.
00:02:12.960 --> 00:02:17.759
So that meant I started my homework at midnight and had maybe two to three hours before it lights out.
00:02:18.000 --> 00:02:21.759
So because of that, I couldn't do digital chores with like organizing whatever.
00:02:21.840 --> 00:02:23.199
Like I just need to get the work done.
00:02:23.360 --> 00:02:25.360
And also everything was always in my head.
00:02:25.520 --> 00:02:29.520
So constantly tracking everything I had to do, juggling between those different things.
00:02:29.759 --> 00:02:36.560
And I think it was around our junior year in college when Kevin and I started working on a little side project.
00:02:36.800 --> 00:02:39.840
People still called everything GPT rappers at the time.
00:02:40.080 --> 00:02:42.400
And I think we were working really late night.
00:02:42.560 --> 00:02:48.800
So we wanted to have GPT 4.0 order DoorDash for us late at night when we forgot to eat dinner.
00:02:48.960 --> 00:02:51.599
And that was the very first like use case for ourselves.
00:02:51.680 --> 00:02:52.960
It was to do that.
00:02:53.120 --> 00:02:53.840
And it worked.
00:02:53.919 --> 00:03:02.319
So then after that, we had it do a variety of other things, like organizing our school work into calendar, then eventually applied it to like an invoicing use case.
00:03:02.479 --> 00:03:07.199
And I think that was the moment when we realized this was maybe like a useful thing for other people too.
00:03:07.439 --> 00:03:10.960
What happened was we snuck into this AI conference in Boston.
00:03:11.280 --> 00:03:12.000
Which one?
00:03:12.240 --> 00:03:13.680
I don't even remember the name.
00:03:13.759 --> 00:03:17.360
And we like found any random event that was happening outside of campus.
00:03:17.599 --> 00:03:24.479
And we had to print out fake lanyards and or name cards and buy lanyards to sneak into the conference too.
00:03:24.639 --> 00:03:32.639
But we had just so happened to bump into someone who he managed, I think, hundreds of like UI Path bots as his job.
00:03:32.800 --> 00:03:38.560
And he saw our demo just casually in the hallway of the of the vet and he was like super surprised.
00:03:38.639 --> 00:03:41.439
And he was telling us like this can replace UI Path.
00:03:41.520 --> 00:03:43.360
Like I have to manage all these computers.
00:03:43.520 --> 00:03:45.280
It always breaks, it doesn't do anything.
00:03:45.439 --> 00:03:51.120
And you guys have this piece of tech that can like just download the PDF and put it into QuickBooks.
00:03:51.439 --> 00:03:54.639
And that was what made us realize, like, hey, maybe this could be useful.
00:03:55.039 --> 00:03:56.719
So you were at Harvard at the time.
00:03:56.800 --> 00:03:58.879
And so did you end up finishing or did you drop out?
00:03:59.199 --> 00:04:01.120
I ended up leaving and dropping out.
00:04:01.199 --> 00:04:01.520
Yeah.
00:04:01.919 --> 00:04:09.280
Did you see any parallels from your years of piano playing to building or coding a new company?
00:04:09.520 --> 00:04:10.400
Yeah, yeah, for sure.
00:04:10.479 --> 00:04:18.240
I think like I was looking back and reflecting, and all those hours I spent really focusing on doing one thing super, super well.
00:04:18.399 --> 00:04:20.319
That definitely like taught me a lot.
00:04:20.399 --> 00:04:32.639
And I think it also led to me pivoting into math as my as my major in in college, because that's concentration where you really have to sit and think about a problem for a really long time.
00:04:32.800 --> 00:04:35.519
Um, and then after that, with building, it's the same thing.
00:04:35.600 --> 00:04:39.199
It's like you have to really commit and be really passionate about it.
00:04:39.360 --> 00:04:46.319
Yeah, like no kid wants to give up all those things repeatedly for something unless they're really, really passionate about that one thing, right?
00:04:46.560 --> 00:04:48.720
It's probably very rewarding for you, also.
00:04:48.800 --> 00:04:53.920
Like if you had all that energy to give to it, there had to be some feedback mechanism.
00:04:54.240 --> 00:05:01.680
Yeah, I I in terms of feedback, I think like the other thing that it taught me was like being really good at taking harsh feedback.
00:05:01.839 --> 00:05:11.360
So, like from piano teachers to public masterclass where you perform and you're getting roasted in front of the the crowd and they're there to take notes and watch you get critiqued.
00:05:11.519 --> 00:05:19.120
And then also competitions where it's very subjective and also you like your whole years of work gets put into one moment on stage.
00:05:19.279 --> 00:05:23.600
But yeah, I do think it's rewarding to have done that for sure.
00:05:23.920 --> 00:05:30.319
Do you think that you're able to show up stronger under pressure now because of your piano experience?
00:05:30.639 --> 00:05:38.240
Yeah, I was just thinking about this yesterday because I was talking, I was catching up with some founders about like high stress situations and running their day-to-day work.
00:05:39.120 --> 00:05:42.240
Yeah, I was thinking about like what is it that makes it stressful?
00:05:42.399 --> 00:05:51.120
And I was thinking for myself, I feel like there's been times where I was so stressed about to get on stage and perform in front of like 100 people.
00:05:51.279 --> 00:05:58.480
And I, for example, didn't have the piece fully memorized because I was doing school and I had maybe two weeks to memorize a 30-minute program.
00:05:58.720 --> 00:06:03.199
Like, what was I gonna do if I like forgot I had just memorized it the day before?
00:06:03.360 --> 00:06:11.600
And that kind of pressure of having to just get on there and fake it till you make it and just hope for the best and and make it out, like having performed something.
00:06:11.680 --> 00:06:18.800
Like I feel like nothing so far has been as daunting as actually doing that, even despite all the difficult decisions to make.
00:06:18.879 --> 00:06:23.920
Like, I know that I can face it head head on after after having done that before.
00:06:24.319 --> 00:06:26.079
There's something about performance, right?
00:06:26.160 --> 00:06:33.279
Like there's showing up in the moment and giving your best, even though you've kind of like prepared and it's never it's never what you imagined.
00:06:33.759 --> 00:06:37.839
Yeah, it never happens the way you expect it to or hope that it would be.
00:06:38.000 --> 00:06:38.240
Yeah.
00:06:38.560 --> 00:06:44.879
Do you remember a time, maybe music or later in life where you really had to like improvise?
00:06:45.360 --> 00:06:46.480
Yeah, yeah, for sure.
00:06:46.639 --> 00:06:49.600
Wait, so was it during music and also after?
00:06:49.839 --> 00:06:55.439
Yeah, one time I actually did forget the page, like the very last two minutes of a concerto.
00:06:55.680 --> 00:06:59.680
It was like the very last whole whole competition piece was like 30 minutes.
00:06:59.839 --> 00:07:02.959
At the very end, I had forgotten it and to improv the ending.
00:07:03.199 --> 00:07:06.879
And I just like made something up and like hit the last note and then walked off the stage.
00:07:06.959 --> 00:07:08.639
Luckily, it was contemporary.
00:07:08.879 --> 00:07:13.680
So when it didn't sound right, like who knows that it might be part of the part of the piece.
00:07:13.920 --> 00:07:14.720
Yeah, definitely.
00:07:14.800 --> 00:07:21.519
I think that has helped me a lot in terms of thinking out of the box when solving problems now.
00:07:21.680 --> 00:07:30.000
Like we operate in a very, I think, like dynamic space where every week something new comes out, you have to think about like first principles like, is what we're doing correct?
00:07:30.160 --> 00:07:31.519
Should we adapt to this right now?
00:07:31.600 --> 00:07:41.040
Or yeah, and and I think that flexibility of being able to handle any kind of situation depending on what's happening in the moment, that's something that's been helpful as well.
00:07:41.360 --> 00:07:44.160
How have you thought about doing that as a first-time founder?
00:07:44.560 --> 00:07:52.879
Like, how are you learning like when to iterate, what hypotheses to test, especially in an exciting but also competitive market that Dex is operating in?
00:07:53.040 --> 00:07:56.560
And it would probably be helpful to touch on Dex as well, just so we can ground folks.
00:07:56.800 --> 00:08:06.560
So Dex is a browser agent that lives inside of Chrome and helps people do manual work in their browser, and it also connects to a variety of integrations outside of the browser.
00:08:06.720 --> 00:08:11.920
So really helping people save time from all the manual admin digital chore work that they have to do.
00:08:12.160 --> 00:08:18.240
And I guess in terms of your question on learning, that's something I've kind of thought about a lot.
00:08:18.319 --> 00:08:25.519
It's like if you haven't seen something before, how do you figure out if you know everything you need to know to make the best decision right?
00:08:25.680 --> 00:08:39.440
And I think like having gone through building for the last year, a lot of times not nothing I can read or learn about from somebody else is as good as just having done it and heard the feedback directly from like your own actions.
00:08:39.679 --> 00:08:48.399
Because I think it's hard to compare situations or get advice because no one fully knows the exact situation and all the details of like your exact product.
00:08:48.559 --> 00:08:52.240
There might be some very tiny detail resulting in some results.
00:08:52.399 --> 00:09:00.080
So, like only you know the full picture, like have the full context on like what might be the actual reason for user feeling a certain way.
00:09:00.240 --> 00:09:11.120
And so I think like it's been a lot more fruitful for us to just do it instead of thinking too much, like trying to just put it out there and then get that feedback from from seeing what actually happens.
00:09:11.440 --> 00:09:13.200
So you describe Dex beautifully.
00:09:13.440 --> 00:09:18.559
Help us ground it a bit in like who is the target for it, information workers or where do they live today?
00:09:18.960 --> 00:09:24.720
So our ICP is someone who fully lives in the browser for like 100% of the day, pretty much.
00:09:24.960 --> 00:09:33.519
And when you think about like where in the industry or like the market those people are, we've tend to find them to be in go to market right now.
00:09:33.679 --> 00:09:42.879
And their entire work day is like in the browser inside of apps and also like a variety of tabs that don't have integrations, like LinkedIn, for example.
00:09:43.039 --> 00:09:51.679
And the other thing that is really interesting is I have been asking everybody I speak to like, do you have a personal to-do list anywhere?
00:09:51.840 --> 00:09:54.080
And for go-to-market people, they never do.
00:09:54.240 --> 00:09:56.879
They don't use the sauna, they don't use linear or anything like that.
00:09:57.120 --> 00:10:03.919
Sometimes their team tries to get them to, but I think the reason is their whole day evolves around other people and it's in the browser.
00:10:04.080 --> 00:10:12.399
So it's way too tedious to spend time putting down their to-dos when their to-dos are literally living like in their email, their calendar, their tabs.
00:10:12.639 --> 00:10:15.759
So that those are the people that we're serving right now.
00:10:16.000 --> 00:10:17.519
I actually spent a lot of time go to market.
00:10:17.759 --> 00:10:20.559
I used to run global operations for a public company.
00:10:20.879 --> 00:10:23.279
Spent a lot of time salespeople and go to market folks in general.
00:10:23.360 --> 00:10:24.720
And many times it's implicit.
00:10:24.799 --> 00:10:27.519
It's like their to-dos are the tabs they have open.
00:10:27.679 --> 00:10:28.000
Yeah.
00:10:28.240 --> 00:10:29.759
The tab being open is the reminder.
00:10:29.840 --> 00:10:32.080
Is the me with blog posts.
00:10:32.320 --> 00:10:35.279
If I open a blog post, they need to read it before the end of the day.
00:10:35.600 --> 00:10:35.759
Yeah.
00:10:35.919 --> 00:10:41.519
They I look at their browser and it's always like four windows maximum to the brim with tabs.
00:10:41.679 --> 00:10:48.879
I think like they try to share a screen with me, and then usually the browser crashes because they're waiting for it to load like every two clicks.
00:10:49.120 --> 00:10:50.320
So we see that a lot.
00:10:50.480 --> 00:10:51.840
And I was actually surprised.
00:10:52.000 --> 00:10:57.200
Like I we thought people had a lot of tabs, but I found people with like way more tabs than I expected.
00:10:57.519 --> 00:11:01.279
What has it been like acquiring early users or going to market?
00:11:01.519 --> 00:11:03.919
How are you finding folks to try ducks out?
00:11:04.320 --> 00:11:07.120
So far, it's been very like natural, word of mouth.
00:11:07.279 --> 00:11:10.799
We haven't invested too much into an actual go-to-market motion yet.
00:11:10.960 --> 00:11:22.240
And this is because we wanted to actually go for like really high retention or make sure we're adding a ton of value into the users who are naturally discovering us first before like growing more with a leaky bucket.
00:11:22.399 --> 00:11:26.399
And I think like as of the last few weeks, we're very happy now with our retention.
00:11:26.480 --> 00:11:28.639
So we'll be doing that more in the future.
00:11:28.879 --> 00:11:36.720
But I think like it's been great to use like LinkedIn and Instagram for us, just having our own team show our own use cases for the product.
00:11:36.879 --> 00:11:38.080
That's actually a big thing for us.
00:11:38.159 --> 00:11:44.159
Like, I'm still like three user of the product today, with like many thousand other people that use it quite a bit too.
00:11:44.399 --> 00:11:47.919
So, yeah, every second of my day, I'm I'm using the product.
00:11:48.240 --> 00:11:54.639
I wanted to go actually back to your point about discovering your your ICP.
00:11:54.799 --> 00:11:58.159
You start in a company, you have this idea, cool.
00:11:58.399 --> 00:12:04.080
One of the characteristics is like almost any human that's in the browser, which is almost any human, can use this.
00:12:04.240 --> 00:12:08.080
How did you end up narrowing it down to go-to-market folks or salespeople?
00:12:08.320 --> 00:12:15.679
Yeah, it's kind of a long story, but before my days building, I w dabbled a little bit in the investment banking world.
00:12:15.840 --> 00:12:22.720
And that was another shocking thing where I looked at Harvard, like, where did the smartest people I know, where were they going?
00:12:22.879 --> 00:12:28.000
And one of them happened to be a pianist, and he was at the time president of every finance club.
00:12:28.080 --> 00:12:29.600
So he was like, you should do this.
00:12:29.759 --> 00:12:31.919
And I was like, sure, this guy did what I did.
00:12:32.000 --> 00:12:33.519
Maybe I should go down that path.
00:12:33.679 --> 00:12:47.360
And during my summer internship and like the time I was kind of recruiting, talking to people in that space, so many very smart, ambitious young people that I knew in college, they left, they entered that industry and they came out just completely burnt out.
00:12:47.440 --> 00:12:53.200
Like they got sent to the hospital, even and every year there's there are people that die like in banking because they're sleeping.
00:12:53.440 --> 00:12:58.720
And then having done the job myself, one of the reasons I actually wanted to explore it was like, what are people doing?
00:12:58.879 --> 00:13:00.639
Like, what's taking that long?
00:13:00.720 --> 00:13:03.440
Like, what why is it like drinking from a fire hose?
00:13:03.600 --> 00:13:14.000
And I got a lot of those questions answered after having seen it, but also like at the same time, a lot of the work that the analysts and people are doing, it's really all automatable by agents today.
00:13:14.240 --> 00:13:23.279
And even I guess like in 2024, it was less possible, but still a lot of it, a lot of it could have been helped greatly tools and better technology.
00:13:23.440 --> 00:13:28.080
So I think I became very passionate right before building decks with that problem.
00:13:28.240 --> 00:13:32.080
I was like, people just should not be having to do this kind of work.
00:13:32.159 --> 00:13:42.080
And even outside of banking, just like regular admin work, there are many people who like every day their job is to do the same like monotonous task in their browser workspace.
00:13:42.240 --> 00:13:50.559
And I just think that if people were freed from doing that coordination work, from moving data between software and all of that, like the world would be a better place.
00:13:50.799 --> 00:13:59.360
So that was kind of what led me to explore where we would be able to kind of like benefit people the most.
00:13:59.600 --> 00:14:08.399
And I think like from there we went from finance to market research when we had applied to YC, which is tied to like one of the other questions.
00:14:08.639 --> 00:14:16.320
And eventually we realized that like we wanted to we started in the browser because that's where we can immediately get the most context.
00:14:16.559 --> 00:14:19.279
So you're able to get a ton of context from the browser.
00:14:19.600 --> 00:14:24.080
What are the capabilities today that you can achieve with that context versus what you can't?
00:14:24.399 --> 00:14:30.879
I think like when most people start out, especially students, they think that being on the desktop is like the best move.
00:14:30.960 --> 00:14:32.639
And this is what we originally did too.
00:14:32.799 --> 00:14:35.519
And very quickly you realize you have a couple paths forward.
00:14:35.600 --> 00:14:40.799
Like either you make a browser extension back before Web MCP came out, and to get more browser.
00:14:41.120 --> 00:14:42.720
You build a whole browser like some have done.
00:14:42.960 --> 00:14:43.279
Yeah.
00:14:43.519 --> 00:14:48.080
Or you have to spend a lot of time working with AppleScript or the Microsoft ecosystem.
00:14:48.320 --> 00:14:50.000
We actually like dabbled in all of those.
00:14:50.080 --> 00:14:53.679
And then we ended up picking the browser because the context is most available.
00:14:53.840 --> 00:14:57.200
There's a lot of tooling provided in the native Chrome ecosystem.
00:14:57.440 --> 00:15:05.600
And beyond just depending on vision, you can also utilize the DOM to have more structured clues into what the user is doing.
00:15:05.759 --> 00:15:07.360
So that was why we picked the browser.
00:15:07.519 --> 00:15:11.759
And I think like now back to the question of like how did we find our current ICP?
00:15:11.919 --> 00:15:17.519
It's really thinking about who depends on the browser as their command center when they work every day.
00:15:17.759 --> 00:15:26.799
And the kind of magic moment early on for our product was just not having to contact switch between their tabs or even just like the organized tabs feature.
00:15:26.960 --> 00:15:32.080
We had spent a lot of time building all these agentic features and like actions that people could do.
00:15:32.240 --> 00:15:41.200
But then at the end of the day, so many people they were like, oh, I just didn't have to switch to this other tab to grab a piece of information, or like, oh, I really like that it cleaned up my tabs for me.
00:15:41.360 --> 00:15:55.600
And that was when we started realizing, like, hey, maybe the value add here is that context switching piece, not like us trying to tell everyone like this can do actions for you, which I think until more recently, people didn't really understand how AI could even do actions for you.
00:15:55.919 --> 00:15:56.480
It's really interesting.
00:15:56.559 --> 00:16:03.039
I wonder if you guys have data on how many times on average people context switch between tasks.
00:16:03.440 --> 00:16:11.840
Yeah, we had this interesting idea to do like a browser wrapped because you can see how productive you were every day, which I think for knowledge workers did never really existed.
00:16:11.919 --> 00:16:18.559
But for engineers, you can check your GitHub history, you can even show people your GitHub, like I did this much engineering work.
00:16:18.720 --> 00:16:26.720
But for with what we've built out, it's actually possible to kind of look back at a diary of your work daily and see like I've been rotating through tabs.
00:16:26.879 --> 00:16:32.240
And what we discovered was like the browser kind of mirrors the mental model of the user.
00:16:32.320 --> 00:16:42.559
And you can tell by matching what their job and their work and their day-to-day looks like to their natural like organization style or mental model around how they organize their day-to-day.
00:16:42.720 --> 00:16:49.200
If you mix and match these things, there's like a finite amount of buckets of types of ways that people kind of like to do work.
00:16:49.360 --> 00:16:53.600
And it's interesting because then you can see like some people rotate through their tabs very quickly.
00:16:53.679 --> 00:16:57.679
Like they linear go through them and they will exit out the moment they stop meeting them.
00:16:57.840 --> 00:17:00.000
And they like cycle through so many tabs every day.
00:17:00.080 --> 00:17:03.279
And then other people leave them all open and they're very attached to the tabs.
00:17:03.519 --> 00:17:10.160
I'm kind of like that, where the the tabs are like reminders for them and they always need to kind of jump to another task, come back.
00:17:10.240 --> 00:17:12.079
So all these threads are just like open there.
00:17:12.240 --> 00:17:19.440
And then you have people who are like very organized by windows, so they'll open up five different windows, and every workflow is in its own little window.
00:17:19.519 --> 00:17:25.039
And I think like Arc users are very much like that because they can organize their spaces easier.
00:17:25.359 --> 00:17:29.759
How do current users evaluate success metrics of the product?
00:17:30.079 --> 00:17:37.440
A successful day of using Dex or like first moment of using Dex is if it was able to take your mind off something.
00:17:37.599 --> 00:17:47.440
So, an example of that, like let's say something as simple as sending an email follow-up or scheduling, you have to pay attention and do that task, even if it's very simple, like responding.
00:17:47.519 --> 00:17:56.160
If your day is already filled back to back with stuff, like you that that extra effort of writing out that response and thinking about it, that is effort for people.
00:17:56.319 --> 00:18:01.759
Um, and that's why they like will switch to GPT, paste it in, and then like have GPT give you a response sometimes.
00:18:01.920 --> 00:18:04.160
They're just kind of tired of that mental effort.
00:18:04.319 --> 00:18:13.359
With Dex, we're trying to increase the difficulty and like the amount of time saved through taking their mind completely off of a task.
00:18:13.440 --> 00:18:17.119
So I can just tell Dex to schedule for me and don't have to think about it anymore.
00:18:17.279 --> 00:18:28.000
Um, they do have to come back to eventually approve some like very mission critical actions, but that whole in-between, like we want them to not have to think about it and just let the agent handle it.
00:18:28.319 --> 00:18:33.359
It's a meta pattern I love to call it offloading your bio tokens into digital tokens.
00:18:33.680 --> 00:18:35.599
Think about how many you're burning a day.
00:18:35.759 --> 00:18:36.000
Yeah.
00:18:36.240 --> 00:18:37.759
These reminders to close your day.
00:18:37.839 --> 00:18:40.640
And if yeah, to your point, easily automatable.
00:18:41.039 --> 00:18:45.279
I feel like 2026 is the year of these browser and computer use agents.
00:18:45.519 --> 00:18:51.200
I think we're seeing the model foundation labs as well as other startups just move deeper into this area.
00:18:51.359 --> 00:18:55.519
How do you think about the landscape and how to quote unquote stay ahead?
00:18:55.839 --> 00:18:57.599
I think a couple of things have changed.
00:18:57.759 --> 00:19:06.319
So, like one, when we started in 2024, fall, computer use had just come out and that became a term like in the very early days.
00:19:06.559 --> 00:19:12.079
And it was very bad, like it was just very slow and it wasn't very accurate.
00:19:12.240 --> 00:19:19.519
And I think that was the case for up until like literally this year in the last few months, it started getting faster and more accurate.
00:19:19.680 --> 00:19:25.599
And I think benchmarks are now around like 80%, and you can further improve that on specific data, right?
00:19:25.759 --> 00:19:37.119
So before everyone was thinking like you can only use this tech really for like full end-to-end automation and like a very traditional, like let's say logistics and go for the UI path route.
00:19:37.200 --> 00:19:41.920
And a lot of people who had been building browser agents actually wanted to use it for consumer.
00:19:42.000 --> 00:19:46.000
And I think also a lot of them were coming from a very engineering heavy perspective.
00:19:46.240 --> 00:19:54.319
Once they actually deployed these, they realized like knowledge workers don't really want to use this in the way that I thought they would want to use this for their work.
00:19:54.559 --> 00:19:56.880
So I think like now it's just that's one piece.
00:19:56.960 --> 00:19:59.759
Like it's a lot more possible, it's a lot faster, it's a lot more accurate.
00:19:59.920 --> 00:20:08.319
And then the other thing is I think there was a very big like mental model shift around going away from this full automation mindset to like co-working is okay.
00:20:08.400 --> 00:20:17.680
And like the whole co-working paradigm is like fine, and having human in the loop is necessary right now and actually improves like the agent's ability to learn and get better too.
00:20:17.839 --> 00:20:26.720
So I think that happened along with the launches of like cowork and the whole claude bot open cloud craze that went very viral.
00:20:26.880 --> 00:20:41.680
And I think those things really helped the people outside of maybe like the SF bubble or people who are less technical to realize like, actually, AI is not just for research and not just for summarizing, but it can do actual actions for you in your like real-world day-to-day tasks.
00:20:41.759 --> 00:20:51.519
And that that's been huge because you can see now everyone's looking for solutions to their problems and they're able to be creative about like what they want AI to actually do for them.
00:20:51.680 --> 00:20:59.119
So, yeah, I think those things combined made browser agents a lot more viable this year and a very exciting space to be in.
00:20:59.440 --> 00:21:04.000
It feels like you know, if we expand on it a bit to your point, like what people are actually trying to solve with this.
00:21:04.720 --> 00:21:11.039
Like almost every day, some playwright CLI came a few weeks ago, and now Alibaba released their page agent.
00:21:11.279 --> 00:21:13.680
And today I saw a couple other things that I can't even recall.
00:21:13.920 --> 00:21:18.160
The ecosystem is exploding, and you guys have been very early on the browser.
00:21:18.480 --> 00:21:23.200
Do you think are you married to the idea of the browser being the right space?
00:21:23.440 --> 00:21:28.079
Or it's more about the purpose, the thing you're trying to resolve and you will adapt to what's happening in the market.
00:21:28.319 --> 00:21:40.799
I think for us, we started on the computer and then we chose the browser because we were super obsessed with like how can we become the player who has captured the most context from the user and truly understand their whole workspace end to end.
00:21:41.039 --> 00:21:42.960
That was the reason why we picked the browser back then.
00:21:43.119 --> 00:21:47.519
There was a lot less standard tooling to your point with like playwright CLI coming out and all that.
00:21:47.599 --> 00:21:56.000
It's showing that like it's becoming more democratized and very standard to be able to access tooling in the browser and that agents need access to the browser.
00:21:56.160 --> 00:21:59.759
So a lot of helpful tooling has come out, but we are definitely not making.
00:22:00.160 --> 00:22:02.079
Married and we're not like browser only.
00:22:02.240 --> 00:22:07.279
Even today, a big misconception that people have is that we're like purely a browser agent.
00:22:07.440 --> 00:22:20.000
And actually, the way that DEX works is we take in context from the browser and then we do actions through the form of tool calling, like not through MCP for us, but we make custom tools for the power apps that we care about.
00:22:20.160 --> 00:22:24.799
And that's been a hundred times more efficient and accurate for our users.
00:22:25.039 --> 00:22:28.400
And we definitely also mixed in browser tooling with that.
00:22:28.480 --> 00:22:37.359
And that was something that we didn't see labs or like people doing that often for quite some time in the last year until more recently is like combining the two together.
00:22:37.519 --> 00:22:47.519
So that means the agent knows to call a browser tool and click on your screen when that's needed, and then it will call like directly into Gmail, for example, when that's more efficient.
00:22:47.839 --> 00:22:52.880
Have you guys seen like Claud and Chrome extension and like Atlas and all those?
00:22:53.039 --> 00:22:53.359
Yeah.
00:22:53.519 --> 00:23:02.880
If you try asking it to fill out a spreadsheet for you, it will take seven minutes plus to try to fill out one row, and then it ends up putting it into one cell.
00:23:03.039 --> 00:23:13.519
And that's just like the most obvious example for people when like you ask Dex to do it, it will be able to fill like hundreds of rows just in a few minutes accurately.
00:23:13.680 --> 00:23:16.640
And I think like the tool calling stuff is very helpful there.
00:23:17.039 --> 00:23:20.880
My agents have been playing a bit with the playwright CLI and the Alibaba agent of payages.
00:23:23.920 --> 00:23:27.519
What do you think are the bigger like constraints in the space right now?
00:23:27.759 --> 00:23:29.920
2026 compared to 2027.
00:23:30.079 --> 00:23:30.880
Where will we go?
00:23:31.200 --> 00:23:37.440
I think right now a lot of attention is on like how to create, gather, and utilize like team context.
00:23:37.519 --> 00:23:48.319
So it's no longer only about like your personal agent, but it's also like how can you share these agents to your team and have this collective data that's been extracted and utilized to make your agents better?
00:23:48.480 --> 00:23:52.799
And it's interesting because I think a lot of people want to build it from scratch right now.
00:23:52.880 --> 00:23:55.359
Like obviously, Cloud Codes made things a lot easier.
00:23:55.519 --> 00:24:05.200
And I think it's like not sustainable long term because it's very like one-off, and you are quite literally building a whole platform end-to-end if you want to make it something your whole team can use.
00:24:05.279 --> 00:24:06.880
Like there's a lot of pieces to learn.
00:24:06.960 --> 00:24:11.200
And when I talk to especially like go-to-market folks, they're already working till like 12.
00:24:11.599 --> 00:24:14.640
Like they have these back-to-back calls, and then they have to learn how to code.
00:24:14.720 --> 00:24:20.160
They have to learn how context works, like how agents work, and then they're trying to vibe code a solution to build agents.
00:24:20.319 --> 00:24:22.319
And it's just a lot, a lot to do.
00:24:22.480 --> 00:24:33.680
And so I think like looking at the year ahead, a lot of the interesting part is around like one, how do you actually get structure the context, retrieve it, and then allow agents to use it?
00:24:33.839 --> 00:24:39.839
And then another big thing we're very focused on is like what's the correct interface or like way to interact with the user.
00:24:39.920 --> 00:24:46.480
I think a lot of it's actually figuring out the right scaffolding around the tech because the tech works now, it's only going to get better.
00:24:46.640 --> 00:24:51.759
But allowing people to actually utilize it in the mental model that they that makes sense to them.
00:24:51.920 --> 00:24:56.000
I think that piece is like under under figured out at the moment.
00:24:56.319 --> 00:25:01.839
So like kind of different people want to use different agents in different places, right?
00:25:01.920 --> 00:25:04.240
And I think you should be able to use it from anywhere.
00:25:04.400 --> 00:25:08.960
So some people might work from the browser, they'll use it from the browser CLI phone.
00:25:09.119 --> 00:25:13.839
Depending on where their day is, like that's where they should be able to use the technology.
00:25:14.160 --> 00:25:17.839
And then there's also a lot of questions with like how do you make sure it's reliable?
00:25:18.000 --> 00:25:22.400
How do you build in observability and all of that is very important to make it truly useful?
00:25:22.960 --> 00:25:28.960
Yeah, as you're building into the into businesses, into the enterprise, obviously trust and governance and yeah.
00:25:29.759 --> 00:25:31.759
How do you guys think about like parallelization?
00:25:32.480 --> 00:25:35.039
Which is notoriously complicated.
00:25:35.279 --> 00:25:38.480
It's hard to drive the browser multiple directions at the same time.
00:25:39.039 --> 00:25:42.559
So we have parallel agents running, calling the tools.
00:25:42.799 --> 00:25:51.680
So for example, if I wanted to reach out with custom personalized emails to this whole spreadsheet of people, we'll send off different agents to go do that task.
00:25:51.839 --> 00:25:55.440
And they just the user just sees these pop up like in their browser sidebar.
00:25:55.759 --> 00:26:02.880
So for difficult tasks like that that are in bulk, we're not sending off browser agents because that will kill their Chrome immediately.
00:26:03.119 --> 00:26:06.240
On the other hand, we play around with the tab groups a lot.
00:26:06.319 --> 00:26:15.359
So we've been trying to see like if we can figure out this tab group or like these scattered tabs are one work stream and one thread, and these other ones are a different one.
00:26:15.440 --> 00:26:27.279
And if we can first of all like tab group things correctly, not by topics of the tab, but by their actual workflow, then can we also have a different sidebar come up to tackle that task in that tab group?
00:26:27.440 --> 00:26:29.039
So this is something we're exploring with.
00:26:29.200 --> 00:26:42.160
As of right now, the way it works is it's one universal sidebar for like all the then people have actually liked that more than some of the AI browsers, where it opens up a new sidebar per tab because then you're still contact switching a lot.
00:26:42.559 --> 00:26:43.279
That makes sense.
00:26:43.519 --> 00:26:45.119
How many people are on the team today?
00:26:45.359 --> 00:26:46.480
It's still very small.
00:26:46.640 --> 00:26:48.480
So we have four people full-time.
00:26:48.640 --> 00:26:51.279
It's my co-ventor and I, and then two engineers.
00:26:51.599 --> 00:26:56.400
How have you thought about building an AI-first product with in an AI-first world?
00:26:56.559 --> 00:26:58.319
Are you able to keep a leaner team?
00:26:58.720 --> 00:27:05.119
Our internal goal right now is actually to run all of our operations on DEX by the end of April.
00:27:05.279 --> 00:27:14.319
So currently we have two growth interns, and their job before they leave is to like automate their own work as much as possible.
00:27:14.480 --> 00:27:20.640
And we've been feeding all of our context to our context engine, which DEX also connects to and runs off of.
00:27:20.799 --> 00:27:23.519
It's been like amazing how much it can do for me.
00:27:23.680 --> 00:27:30.799
So, like to give you a tangible example, I was evaluating CRMs in the last little bit because I'm now talking to too many different customers.
00:27:30.960 --> 00:27:34.400
They're telling me different feature requests and they all have different blockers.
00:27:34.640 --> 00:27:45.839
And I couldn't find really a CRM that's far enough where it can actually gather all of my data from everywhere and then surface the right results with like AI to analyze and consolidate that.
00:27:46.079 --> 00:27:57.519
And then after we had connected our contract system to all these integrations that we have, so it's live updating with every Slack message, every email I'm getting, every call, calls that I'm not in as well.
00:27:57.680 --> 00:27:59.519
And all of that's feeding in live.
00:27:59.599 --> 00:28:05.279
And at any moment, Dex has access to that and knows and has this complete understanding of what's going on.
00:28:05.519 --> 00:28:11.759
So basically, what I asked it to do was just like generate me a spreadsheet with all the fields that I care about.
00:28:11.839 --> 00:28:16.240
Like I want to know what's blocking every customer right now that I spoke to in the last week.
00:28:16.400 --> 00:28:21.839
I also want to know what feature requests have been most requested and what should be prioritized in our product roadmap.
00:28:21.920 --> 00:28:24.559
And it was able to do that in I could do that like one minute.
00:28:24.640 --> 00:28:27.519
And it was more comprehensive than any CRM that I tested.
00:28:27.680 --> 00:28:30.799
And I can just whip up an analysis like that at any given moment.
00:28:30.960 --> 00:28:32.559
So that's been super powerful.
00:28:32.720 --> 00:28:34.960
And that's not even an automation use case.
00:28:35.039 --> 00:28:40.799
Like I can have this run on a trigger, I can do a bunch of things to tie like all the agents in the background together.
00:28:40.880 --> 00:28:42.960
So yeah, it's definitely very powerful.
00:28:43.200 --> 00:28:46.079
And we are planning to run our company off of that.
00:28:46.319 --> 00:28:47.839
That sounds a lot like a memory problem.
00:28:48.000 --> 00:28:52.559
That feeds into your context engine, or are you guys using something from the community for that?
00:28:52.960 --> 00:28:54.720
We build everything from scratch.
00:28:54.880 --> 00:28:58.319
So, how the way that we handle context and memory, there's two pieces.
00:28:58.559 --> 00:29:13.920
One is kind of the source of truth that we're extracting from integrations, and that just comes from having a lot of connectors, extracting the data out, and building relationships and understanding what is this company's source of truth or like their reality, essentially.
00:29:14.079 --> 00:29:20.960
And then the other piece comes from observing in the browser what the daily processes and decisions are being made coming from each individual.
00:29:21.119 --> 00:29:27.519
And we combine these two together to actually like enrich the knowledge of what Dex can understand and know.
00:29:27.680 --> 00:29:34.400
And the two pieces kind of give you different pros and cons, but it builds a much better understanding of everything.
00:29:34.799 --> 00:29:40.799
How far do you think you'll be able to scale as a company with call it four, five, six people?
00:29:41.599 --> 00:29:42.960
I think you can do a lot.
00:29:43.119 --> 00:29:48.319
Like what I realized is that software is easy to make now.
00:29:48.480 --> 00:29:57.759
Like the like the CRM piece or like a dashboard, or if you have full access to your data and you can utilize it with agents, it can do so much for you.
00:29:57.839 --> 00:29:59.599
It can just, you can just vibe code a CR.
00:29:59.839 --> 00:30:07.519
If you wanted one, you can vibe code a dashboard or an automation or an agent, and all of it can kind of do work for you when you're resting and sleeping.
00:30:07.839 --> 00:30:11.039
So I think like that's the direction I see things going.
00:30:11.119 --> 00:30:13.680
It's almost like lovable, but for agents.
00:30:13.839 --> 00:30:20.160
And what needs to happen though is like you need to be able to extract and understand all the data that you have.
00:30:20.240 --> 00:30:36.960
And then there's a missing piece of data, which is what we're trying to figure out how to collect more of, which is kind of like how are you people making decisions and what actually results in a certain outcome that gets logged somewhere versus just looking at the outcome and not knowing how that got there.
00:30:37.359 --> 00:30:40.720
Building on that, you're a very lean team, you're achieving a lot, you're moving really fast.
00:30:40.960 --> 00:30:50.319
Is there some magic sauce in how you organize your engineering best practices or things that you've developed along the way to help you to do that?
00:30:50.640 --> 00:30:50.799
Yeah.
00:30:50.960 --> 00:30:57.519
So with building off of what I said earlier with the CRM that I can just create at any second, I have that update.
00:30:57.680 --> 00:31:01.759
Every time I have new conversations after a few days, I'll ask it to update that.
00:31:01.920 --> 00:31:08.240
And then with that, I immediately tell it to prioritize like the product roadmap items we have and add it all to our linear.
00:31:08.400 --> 00:31:13.200
So that piece, no one has to kind of manually add tickets into linear.
00:31:13.440 --> 00:31:21.200
Anytime anyone gets like a customer request, either through a call, an email, or in Slack, we also just ask Dex to go add it into Linear.
00:31:21.440 --> 00:31:23.920
So all of that work saves so much time for us.
00:31:24.079 --> 00:31:27.519
And you can also, that's like the day-to-day piece.
00:31:27.680 --> 00:31:32.160
You can also kind of take a step back and it helps me analyze like the direction that we're going in.
00:31:32.319 --> 00:31:40.160
So I use it a lot for brainstorming and looking at all the data at once, which is not something you can humanly do by yourself.
00:31:40.240 --> 00:31:43.440
Like that's I think one of the greatest use cases for AI.
00:31:43.680 --> 00:31:56.079
I can have it read through every single interaction that we've had in the last few months, or like look through logs and figure out, like on the broader scheme of things, what are we getting wrong in our assumptions of our users?
00:31:56.240 --> 00:32:00.799
Like, is is what we think correct, or is there something else that's going on here?
00:32:00.960 --> 00:32:05.440
So constantly like analyzing and reanalyzing and figuring it that out.
00:32:05.599 --> 00:32:08.640
And I think that's something I would like very much recommend as well.
00:32:08.960 --> 00:32:16.319
Maybe we could do one last subject around where Dex is today, like what's keeping you up at night?
00:32:16.880 --> 00:32:27.119
I think one big thing that's a little bit surprising, and actually it's not that surprising once you think about it, is that one of the blockers at the end of the day is human creativity.
00:32:27.279 --> 00:32:36.640
Like people are not very good at systems thinking or managing others, and giving them agents is essentially telling them you have this team that you can go tell them to do whatever.
00:32:36.880 --> 00:32:41.759
And what we found is that people find it hard to come up with new use cases.
00:32:41.839 --> 00:32:52.480
So even if we give them templates or we give them suggestions, sometimes they're like, oh great, I can use decks for this one use case, and they'll always use it for that one use case, but they're not able to think of more use cases.
00:32:52.640 --> 00:32:54.559
So that was like a big learning for us.
00:32:54.880 --> 00:33:01.039
It's sometimes better to educate your user on the capability rather than a feature.
00:33:01.200 --> 00:33:04.640
And a lot of times people will take use cases as features.
00:33:04.720 --> 00:33:13.279
They're like, okay, I can do this use case, but that to them that's a feature rather than like, I understand the capabilities underneath and I can now apply it to other situations.
00:33:13.519 --> 00:33:26.079
So we've been working a lot on kind of like the education piece, and I think that's actually a plus when it comes to sales because many teams right now, they're also internally trying to figure out is the right move to adopt a product?
00:33:26.240 --> 00:33:33.759
Is it to build internal tooling, or should we actually invest in like a cloud code program to teach everyone how to use AI and how to build agents?
00:33:33.920 --> 00:33:43.599
I guess like education is a big piece of what we're doing, and we're like contributing to helping people actually understand how to use agents and like what's going on underneath the surface.
00:33:43.920 --> 00:33:46.319
That's such a beautiful, you've crystallized it very well.
00:33:46.480 --> 00:33:52.799
I find that AI is fundamentally very new and unlocked a lot of things, but also in many ways it's not new at all.
00:33:53.039 --> 00:33:56.559
To your point, the limiting factor seems to be thinking and levers.
00:33:56.799 --> 00:34:02.480
Humans who can understand the levers of a system have found ways historically how to achieve leverage, right?
00:34:02.559 --> 00:34:04.559
Hiring other people, building systems.
00:34:04.720 --> 00:34:07.599
Now it's kind of like AI is bringing leverage to anyone.
00:34:07.759 --> 00:34:07.920
Yeah.
00:34:08.079 --> 00:34:12.320
And if you don't have that thinking, you're just gonna phase zero interact with it back and forth.
00:34:12.400 --> 00:34:14.639
But if you think in a system, you can do so much.
00:34:14.960 --> 00:34:18.559
Yeah, and that's also why I brought up kind of like the lovable point, right?
00:34:18.639 --> 00:34:22.239
Because when websites first came out, it was difficult to make one.
00:34:22.320 --> 00:34:25.519
And everyone wanted to prove the fact they could build it from scratch.
00:34:25.599 --> 00:34:31.119
And then more and more tooling came out, and now we have like you can vibe code a full stack app or a website super easily.
00:34:31.199 --> 00:34:44.880
So I think the same thing will eventually happen with like building systems and agents, hopefully, where you don't have to think about as much going on behind the scenes and like the tooling that we're building and other people are building kind of take care of a lot of that process for you.
00:34:45.039 --> 00:35:00.000
So, like everything from context management to prompting correctly to routing to the right type of agent to use, like whether that's browser or like tool calling or all of that is kind of knowledge that it's important to have that knowledge to know how to optimize your workflow.
00:35:00.159 --> 00:35:08.960
And I realized that the reason why some users aren't able to use it as much as I can is because they don't have that knowledge and they wouldn't know that.
00:35:09.119 --> 00:35:10.159
Why would they know that?
00:35:10.320 --> 00:35:13.599
That's like something that will just improve gradually throughout the year.
00:35:13.920 --> 00:35:14.480
Meta, right?
00:35:14.639 --> 00:35:21.199
Like goes all the way back to that advice you got, that school contact that led you down the path to investment banking.
00:35:21.360 --> 00:35:25.199
And only through the pain you realize what you want to do and you understand the system.
00:35:25.280 --> 00:35:27.599
And now you built like a system of leverage all around that.
00:35:27.760 --> 00:35:28.159
Yeah.
00:35:28.400 --> 00:35:30.239
You have to go through the pain to learn.
00:35:30.480 --> 00:35:39.360
And that's actually also we didn't touch on this, but why we pivoted inside of our YC batch, we just realized that there was nothing wrong with the idea we were doing.
00:35:39.599 --> 00:35:40.159
What was it?
00:35:40.400 --> 00:35:42.719
We joined YC building customer intelligence.
00:35:42.800 --> 00:35:46.000
So this was around when we were really obsessing over context.
00:35:46.079 --> 00:36:00.320
And then we had this idea that what if you could have your product collect user touch points directly and use the interface to interact with users, collect those that data, and utilize that to adapt your website or adapt your product.
00:36:00.559 --> 00:36:05.599
And the reason why we moved out was purely kind of like a founder product fit reason.
00:36:05.760 --> 00:36:15.840
Like we wanted to build the original side project we were building back at school and not something because we thought like this is what people wanted or like it's B2B and all these other reasons.
00:36:16.000 --> 00:36:23.280
I think I actually had like a 30 to 50% response rate when I was messaging CMOs about the like customer market intelligence tool.
00:36:23.440 --> 00:36:35.199
But even that was not like as interesting to us than finding something that was kind of a mix between getting to build at the forefront of the tech, but also something that immediately people, ourselves and people around us like wanted to actually use.
00:36:35.360 --> 00:36:38.960
I think that's pretty rare of an idea to find and have that timing.
00:36:39.119 --> 00:36:42.000
So we couldn't really, we didn't want to lose that opportunity.
00:36:42.320 --> 00:36:42.719
For sure.
00:36:42.880 --> 00:36:46.480
I mean, it's really impressive that you pivoted and it's a long journey.
00:36:46.639 --> 00:36:51.280
So I I think it makes sense to do something that you're passionate about, especially with people that you're so close to.
00:36:51.440 --> 00:36:54.880
I bet you and your co-founder have such an authentic open relationship.
00:36:55.280 --> 00:37:04.000
I can't imagine not knowing, like having to work with someone where you don't know them for like very long, because then you never know what they're really like thinking.
00:37:04.079 --> 00:37:09.599
Whereas, like with my co-founder, we've been friends for so long, we've worked in many different situations for so long.
00:37:09.760 --> 00:37:15.599
Like, if he walks in the room after talking to a potential hire customer, I immediately get the vibe of what's going on.
00:37:15.760 --> 00:37:17.920
And that's like super nice to be able to have.
00:37:18.320 --> 00:37:21.280
Do you and your team you do you guys all live in the same house too?
00:37:21.599 --> 00:37:24.239
We used to all live together as well with our early employees.
00:37:25.840 --> 00:37:27.440
I think that's a very San Francisco thing.
00:37:27.519 --> 00:37:30.960
I spend my time between San Francisco and Tel Aviv that is not happening in Tel Aviv.
00:37:31.280 --> 00:37:33.599
Yeah, it's a little less frequent in New York, I would say.
00:37:34.320 --> 00:37:36.719
It's definitely an interesting experience.
00:37:36.960 --> 00:37:39.840
We all now moved out to be right across the street.
00:37:39.920 --> 00:37:41.840
So it's a 10-second walk for all of us.
00:37:42.000 --> 00:37:44.480
And I think it does make it evolving.
00:37:44.639 --> 00:37:44.800
Okay.
00:37:45.039 --> 00:37:46.719
We have our own doors now.
00:37:47.039 --> 00:37:49.840
Yeah, and we needed more meeting space in our office.
00:37:50.000 --> 00:37:54.480
So it's it's been nice to get like that sunlight when you walk to work in the morning.
00:37:54.800 --> 00:37:56.960
Sunlight seems to be important for humans.
00:37:57.199 --> 00:37:58.239
Yeah, for sure.
00:37:58.480 --> 00:38:00.480
Well, do you want to wrap with a quick fire round?
00:38:00.639 --> 00:38:01.199
Yeah, sure.
00:38:01.360 --> 00:38:01.599
Okay.
00:38:01.760 --> 00:38:05.039
What is one daily task you'll never do again in two years?
00:38:05.280 --> 00:38:08.960
Because it's two years, I'm gonna say something kind of extreme.
00:38:09.199 --> 00:38:14.559
I think by in in two years, I don't want to open a web app and use it like at all.
00:38:14.800 --> 00:38:18.159
I already like have gotten rid of a lot of that in my day-to-day.
00:38:18.320 --> 00:38:20.880
So I don't have to manually respond to emails.
00:38:20.960 --> 00:38:26.320
I don't have to schedule, like even opening Slack, updating linear, all of that's been like done through through Dex.
00:38:26.559 --> 00:38:31.119
And I think like in in two years, I wouldn't really need to like manually use the browser at all.
00:38:31.440 --> 00:38:35.119
What's the most underrated consumer AI use case right now?
00:38:35.440 --> 00:38:40.480
I think it'd be actually reminding you to do something or reminding you not to do something.
00:38:40.559 --> 00:38:43.920
So kind of like proactive, like keeping you focused.
00:38:44.079 --> 00:38:46.480
That's been something that's been really helpful and fun.
00:38:46.639 --> 00:38:48.159
And it's also a little bit gamified.
00:38:48.320 --> 00:38:58.880
Like if you have trouble with getting distracted, it can let you know, like, hey, you need to focus on whatever you're working on, or like the opposite, where maybe you forgot to put something into your system of record.
00:38:58.960 --> 00:39:05.760
Like you wouldn't really know that until like now you have this agent that can tell you you forgot to follow up or there's a task that is about to be due.
00:39:05.920 --> 00:39:08.480
These have been really like fun, interesting use cases.
00:39:08.800 --> 00:39:11.920
What's a non-obvious constraint in building agents?
00:39:12.239 --> 00:39:14.719
Definitely the interface piece of it.
00:39:14.800 --> 00:39:23.199
I think this is under talked about because it's always about like the technicalities of it, but actually like the interface itself, I think is really important.
00:39:23.360 --> 00:39:30.159
And it's because agents fundamentally operate around like your mental model and that includes like the interface you're using it out of.
00:39:30.239 --> 00:39:36.880
And when the CLI or like the IDE is not your natural interface, what should it look like?
00:39:36.960 --> 00:39:40.960
It can take many different form factors, you can interact with it in many different ways.
00:39:41.199 --> 00:39:44.320
Last one, what are you personally too stubborn to let AI do?
00:39:44.639 --> 00:39:50.639
Writing, like everything I write, I still will always make like a first draft of it or speak.
00:39:50.800 --> 00:40:02.880
I I will actually write out like my raw thoughts, even if it's not organized or cohesive, and then I will use AI to come up with more ideas to phrase things, but I don't want it to come up with the first copy for anything.
00:40:03.199 --> 00:40:13.039
The other thing is, oh, I another one that I'm huge, I'm very like, I think like AI interviews or anything where it replaces like you actually talking to someone face to face.
00:40:13.199 --> 00:40:15.679
There's just so much nuance that would get lost.
00:40:15.760 --> 00:40:18.719
Like, I could not imagine ever letting that go.
00:40:18.960 --> 00:40:19.679
I agree.
00:40:19.840 --> 00:40:20.800
Thanks, Regina.
00:40:21.280 --> 00:40:22.239
Thank you guys so much.
00:40:22.400 --> 00:40:23.039
Yeah.
