S13 Bonus: The Spatial Web Shift: Building an AI-Native Browser supporting Agentic AI with Charles Maddock, Strawberry Browser 的完整文字稿

face and since then we have just kept working towards 04:40.000 --> 04:51.568 [SPEAKER_02]: this dream of building the most intuitive interface for non-technical teams doing sales, marketing, talent, and they want to bring agents into their team right. 04:51.588 --> 05:03.176 [SPEAKER_02]: And then just looking forward into the future, we believe that the gigantic web, this new internet that's starting to arise, well, you're going to have trillions of agents communicating with each other. 05:03.756 --> 05:05.097 [SPEAKER_02]: We think this is potentially 05:06.078 --> 05:20.791 [SPEAKER_02]: These most valuable space right now in AI and by building a browser where positioning ourselves to be at the core of that and be the interface that the world's company is used to interface with this new agente web. 05:23.093 --> 05:30.460 [SPEAKER_03]: Tell me about the MVP so that first version you created, how would I take to build and what sort of tools we're using to bring it to life? 05:32.137 --> 05:38.184 [SPEAKER_02]: It's funny story actually, so before we decided to build a browser, we were building a SDK. 05:38.284 --> 05:42.149 [SPEAKER_02]: So a developer told that allowed engineers to build their own web agents. 05:42.690 --> 05:46.855 [SPEAKER_02]: So we're still excited about web agents and the web, but we had a slightly different approach. 05:47.723 --> 05:51.967 [SPEAKER_02]: But we started to realize that DevTools is hard to build for developers. 05:51.987 --> 05:52.827 [SPEAKER_02]: They're very picky. 05:53.067 --> 06:00.093 [SPEAKER_02]: They always want to have very unique features for their cases and we wanted everyone in the world they able to use our product. 06:00.814 --> 06:02.636 [SPEAKER_02]: So just started to become more and more clear. 06:02.876 --> 06:06.259 [SPEAKER_02]: We're building the tools for people to build web agents. 06:06.599 --> 06:08.601 [SPEAKER_02]: Why don't we just create a web agent? 06:09.181 --> 06:12.085 [SPEAKER_02]: that everyone can use regardless of technical level. 06:12.666 --> 06:17.814 [SPEAKER_02]: And then slowly but surely the idea started to come up okay what if we just feel the browser. 06:18.708 --> 06:24.652 [SPEAKER_02]: But that's like a pretty insane thing to say in 2025, like who builds a browser in 2025, right? 06:24.792 --> 06:34.858 [SPEAKER_02]: People call this stupid, we're doing that, but in the end, my co-founder Sebastian, it was so convinced that it was the right thing to do, that he just showed up to the office one day. 06:35.399 --> 06:43.824 [SPEAKER_02]: I was like, guys, I did it, I did it myself, and then over the weekend he had just used a cursor at the time. 06:43.844 --> 06:44.985 [SPEAKER_02]: That was the earliest. 06:45.725 --> 07:03.838 [SPEAKER_02]: like really good vibe coding tool to build this MVP and it was amazing so that that was the earliest MVP right it was a super simple browser with a sidebar where you could ask an agent hey could you click on this page could you feel this former could you help me answer this question right 07:05.219 --> 07:13.681 [SPEAKER_03]: Let's stay on that MVP for a second and with my next question, I could probably cherry pick some answers from what you said, but I'm curious what you would consider. 07:14.321 --> 07:19.582 [SPEAKER_03]: A core decision you had to make and building that MVP and I could hear something around keeping it simple. 07:19.922 --> 07:25.524 [SPEAKER_03]: The tools that were used and things like that may be acceptance of focus on it at a very specific way. 07:25.844 --> 07:30.845 [SPEAKER_03]: I'm curious about what you think was the core decision though and how you cope with making that decision. 07:32.363 --> 07:46.609 [SPEAKER_02]: One of the things that set as a part quite a bit from the other AI browsers early on at least was that we actually made the boring decision to make it look more similar to Chrome than the rest of the competition. 07:47.070 --> 07:51.812 [SPEAKER_02]: The reason for that was stemming from this idea that we want to make it as intuitive as possible. 07:52.572 --> 07:55.233 [SPEAKER_02]: I'll give you a concrete example at last, for instance, the 07:59.595 --> 08:05.057 [SPEAKER_02]: That one had, when you opened a new tab, it focused on the chat input. 08:05.738 --> 08:12.400 [SPEAKER_02]: So it assumed that the user wanted to send a task to their agent, rather than the address bar rights. 08:13.140 --> 08:19.103 [SPEAKER_02]: And most of the AI browsers were like that, they assumed that you would want to chat with the agent since it's a new paradigm. 08:19.543 --> 08:25.005 [SPEAKER_02]: We said, hey, let's make this look more like Chrome and just have the normal features, it's more boring. 08:25.365 --> 08:26.666 [SPEAKER_02]: But what we noticed was that, 08:27.306 --> 08:44.422 [SPEAKER_02]: people really appreciated just being able to move their stuff over from chrome and have the same behavior, you can use the same shortcuts, but then you have this helpful AI companions that could slowly start integrating more and more into your work, so you don't get shell shocked when you. 08:45.022 --> 08:47.224 [SPEAKER_02]: download the new browser and expect it to be different. 08:48.024 --> 08:55.309 [SPEAKER_02]: That's probably one of the decisions we made that we felt bad about making because you always want to innovate right. 08:55.589 --> 09:01.133 [SPEAKER_02]: But I think it was the right choice to make it more familiar and intuitive to the majority of our users. 09:04.127 --> 09:05.108 [SPEAKER_03]: Let's move forward then. 09:05.308 --> 09:09.413 [SPEAKER_03]: So, how have you progressed and mature the platform? 09:09.453 --> 09:12.677 [SPEAKER_03]: I know a lot of that's going on right now as well. 09:12.837 --> 09:15.480 [SPEAKER_03]: So it makes me curious about how do you build your roadmap? 09:15.760 --> 09:17.262 [SPEAKER_03]: How do you go about deciding that? 09:17.342 --> 09:22.568 [SPEAKER_03]: Okay,

this is the next most important thing to build or to address or to shape the strawberry. 09:23.892 --> 09:50.624 [SPEAKER_02]: It's funny because I would say the people that we attract and bring to strawberry are these super passionate builders and they've ran startups in the past usually built their own products with very, very highly in our recruiting process, so it's quite easily becomes this kind of wild west situation where everyone has ideas and we usually people have very good ideas. 09:51.724 --> 10:04.296 [SPEAKER_02]: It becomes a situation where a lot of things want to get built and of course being able to use agents to code makes that situation even crazier because you can build so many things. 10:04.956 --> 10:08.199 [SPEAKER_02]: So I would say in the beginning the way we worked was quite 10:09.453 --> 10:09.673 [SPEAKER_02]: loose. 10:09.693 --> 10:18.540 [SPEAKER_02]: We had some ideas we spoke with users on what we wanted to build, but then we were also constantly thinking, what do people that use the browser every single day do? 10:18.960 --> 10:22.443 [SPEAKER_02]: If they're non-technical, right, they're in meetings, okay, we should probably have a meeting tool. 10:22.943 --> 10:29.608 [SPEAKER_02]: They're doing research, and creating pitch decks, all then we should build a really good research agent and also allow to generate pitch decks. 10:30.429 --> 10:43.536 [SPEAKER_02]: So I would say it's very free for all when it comes to ideas and then we have all hands where we discuss the top things that we need to prioritize and it comes slightly from user feedback and also actual calls with customers. 10:44.417 --> 10:46.378 [SPEAKER_02]: But I would definitely say as we grow, 10:47.118 --> 10:51.900 [SPEAKER_02]: we should find better structures for prioritizing what to work on. 10:52.320 --> 11:04.884 [SPEAKER_02]: But yeah, just constantly working towards building the most delightful agent that form for non-technical users to essentially start bringing agents into their lives and their work. 11:07.025 --> 11:12.587 [SPEAKER_03]: Okay and in what you're describing here is saying, we tell me about how you built your team and what you look for in those people 11:17.072 --> 11:24.935 [SPEAKER_02]: even in the age of agents when you have so many agents working for you, the people on your team are your most valuable asset. 11:25.936 --> 11:31.698 [SPEAKER_02]: We have a couple of things we value extremely highly, we value high rationality, 11:32.367 --> 11:34.287 [SPEAKER_02]: low ego characteristics, right? 11:34.548 --> 11:42.890 [SPEAKER_02]: So people who are essentially ready to give up something that worked on for two weeks, if it's the wrong direction, right? 11:43.570 --> 11:56.853 [SPEAKER_02]: You don't want to be in a situation where someone on the team was working on the wrong thing, but couple their ego to that feature so that when you as a team sit down and you say, hey, this is completely the wrong focus, 11:57.513 --> 12:19.761 [SPEAKER_02]: then they will start protecting that feature and then you get the situation where the organization just moves in the wrong direction however if that same person has such a low ego that they can say yeah you know what you guys might derite let's put this to the side you can move faster together we also value previous high agency 12:20.541 --> 12:21.722 [SPEAKER_02]: projects or situations. 12:22.202 --> 12:43.575 [SPEAKER_02]: So if we can see that someone has worked on a start-up or a project and they actually managed to sell it or it was state-of-the-art within some technology and receive some kind of concrete appraisal, whether it's GitHub stars or followers, I think that's extremely important. 12:43.775 --> 12:47.477 [SPEAKER_02]: And with much, much higher than any kind of academic 12:49.738 --> 12:55.000 [SPEAKER_02]: And then finally, we always do a work trial, and this is so important. 12:55.400 --> 13:10.106 [SPEAKER_02]: We always have a candidate when we're hiring, come over at least one full day, so we can sit together, have lunch, get a feeling for how the cultural fit is, and then during that day, we give them a really difficult technical challenge. 13:10.286 --> 13:12.507 [SPEAKER_02]: So that that's the fourth and final thing. 13:12.627 --> 13:16.909 [SPEAKER_02]: Of course, the person in question has to be really sharp developer. 13:17.730 --> 13:20.436 [SPEAKER_02]: And they can't just let agents code everything right then. 13:20.897 --> 13:27.312 [SPEAKER_02]: Still need to explain their thought process and be able to make changes manually, even though we have this agent right. 13:31.559 --> 13:53.663 [SPEAKER_03]: I'm curious about scale and where scale comes into play here when you're building for something like what you're building for was strawberry browser and where you want to go with it and agentic AI I'm curious where scale has to be thought of and if you thought of it in a certain way in the beginning but I'm also interested if there's been interesting areas where you've had to fight scale as you've grown and that could be people or technology. 13:55.370 --> 14:03.857 [SPEAKER_02]: I'd say one interesting thing with thought a lot about is because web agents are the most expensive class of agent to run, right? 14:03.937 --> 14:16.508 [SPEAKER_02]: They're not only looking usually at large amounts of markedown or HTML in each step, but they're also looking at screenshots and might take a few iterations to submit a forum or use a page. 14:17.569 --> 14:18.810 [SPEAKER_02]: They easily stack up 14:19.552 --> 14:21.154 [SPEAKER_02]: a large bill when they run. 14:22.216 --> 14:23.157 [SPEAKER_02]: So how do you counter that? 14:23.217 --> 14:34.371 [SPEAKER_02]: Of course, you can optimize the harness to work better at scale, but that's actually what a lot of our competitors did, like Atlas, like comet, and the result is just a worse agent experience. 14:35.052 --> 14:37.694 [SPEAKER_02]: So, how

did we solve that scaling issue? 14:37.914 --> 14:43.317 [SPEAKER_02]: We just try to make it as clear as possible in our messaging and in our product that will. 14:43.917 --> 14:54.183 [SPEAKER_02]: This is an expensive tool and you need to figure out the return on the investment and your companion is also going to help you figure out the return on the investment. 14:54.223 --> 14:55.243 [SPEAKER_02]: So let me give you an example. 14:55.784 --> 14:57.845 [SPEAKER_02]: You could use strawberry. 14:58.505 --> 15:01.006 [SPEAKER_02]: to find a new knife for your kitchen, right? 15:01.446 --> 15:06.908 [SPEAKER_02]: And it's going to go into Amazon, it's going to click around, it's going to cost a bunch of credits, and ultimately, maybe you save 10 bucks. 15:07.608 --> 15:16.832 [SPEAKER_02]: But you could use the same amount of credits with the same agent harness to find a founding engineer for your team that is a perfect fit, right? 15:17.772 --> 15:25.840 [SPEAKER_02]: And of course, when I do that, instead, and it costs the same amount of tokens, then I've potentially saved thousands of dollars, right? 15:26.401 --> 15:39.394 [SPEAKER_02]: So we need to think about the return on investment that's been extra important for us, but I think this kind of cost calculation can be done by really, any founder doing building agents, and then thinking about scaling the team. 15:39.995 --> 15:50.711 [SPEAKER_02]: We have a very high bar for the tin and we have we religiously follow a principle called the keeper's test. 15:51.476 --> 16:13.714 [SPEAKER_02]: it is if executed in the right way one of the best ways to build a really strong team whilst not being too mean against candidates and basically it was developed by Netflix and you do a interview process you figure out if someone is good enough to start or not and then after that point. 16:14.334 --> 16:31.589 [SPEAKER_02]: You add 30 days, 60 days, and then 90 days, ask yourself the question, if this person was going to get approached by a different company and they wanted to try to coach them away from my company, how hard would I fight to keep that person on my team? 16:32.210 --> 16:41.458 [SPEAKER_02]: And you ask yourself that question and you're super honest with yourself and if the answer is, it probably wouldn't be too bad if this person got approached. 16:42.118 --> 16:56.183 [SPEAKER_02]: Then, that person, and this will sound very harsh, but that person probably should not be on your team in the first place if you want to keep the talent very high and not skillly amount of people in your team too fast. 16:57.103 --> 17:03.705 [SPEAKER_02]: And of course, if you feel like, okay, say that you were on my team Noah and someone tried to approach you and I was like, oh my goodness. 17:04.005 --> 17:08.227 [SPEAKER_02]: If Noah gets poached, the whole company is going to get brought to its knees well. 17:09.307 --> 17:24.935 [SPEAKER_02]: Then of course that's an engineer that you need to keep and then you have one on once every 30, 60 and 90 days and this has helped us keep the line and the level extremely high and I recommend this to every founder I speak with. 17:27.056 --> 17:30.898 [SPEAKER_03]: I'm curious about as you step on the balcony and you look across all that you 17:35.929 --> 17:41.752 [SPEAKER_02]: we've built a lot of stuff that I would consider state of the art and new ways of using agents. 17:42.752 --> 17:49.755 [SPEAKER_02]: But the thing I'm the most happy and satisfied about is the approach we took to representing the agents. 17:50.736 --> 17:57.299 [SPEAKER_02]: So I don't know if you saw strawberry but it contains this small, cute, colorful companions. 17:58.059 --> 18:10.797 [SPEAKER_02]: and it's very difficult, you really need to strike the balance between a playfulness but not too childish, because if they're too childish then they're not going to get serious adoption in enterprise, right? 18:11.538 --> 18:12.259 [SPEAKER_02]: And I think 18:13.100 --> 18:29.581 [SPEAKER_02]: The balance we found now where they feel smart, elegant, hard working without being too childish or too sterile has been something we really tried to perfect and we've got it to a nice place where we see teams that consist of 50 or 50 year olds that are 18:30.422 --> 18:51.658 [SPEAKER_02]: never used AI, don't know care about AI, but they have created their own companion, they shared with the entire team and they are like printing t-shirts with their mascot on it and they're like, I'm addicted to strawberry, so I think the embodiment that we managed to create for this agent has set us apart a lot and made it much easier to to let teams 18:52.318 --> 19:04.904 [SPEAKER_02]: adopt AI and actually understand that this is not just a prompt box, this is a colleague that learns over time, it has its place in your team and it has different relationships with different people on the team. 19:07.165 --> 19:08.606 [SPEAKER_03]: Let's flip the script a little bit. 19:08.686 --> 19:10.647 [SPEAKER_03]: Tell me about a mistake you made and how you 19:13.965 --> 19:29.775 [SPEAKER_02]: As a team of super passionate builders, there has definitely been a lot of times where we worked considerable amount of time on a feature, just to realize that when we actually put it out into production, it wasn't the way we expected it to be. 19:30.556 --> 19:34.618 [SPEAKER_02]: Users didn't really understand how it works, a concrete example of this was, 19:35.599 --> 19:44.942 [SPEAKER_02]: Almost a year ago we had a feature where you could record your screen inside strawberry to show how a process was done to one of your companions. 19:45.862 --> 19:56.125 [SPEAKER_02]: And we thought this would be like the next way people would create skills for their agents, but it turns out that at least at the time people were very bad at. 19:57.165 --> 20:09.753 [SPEAKER_02]: explaining as they went along what they wanted to do and the end result was usually just at the agent

received a long sequence of video with kind of vague descriptions that it couldn't really do anything with. 20:10.653 --> 20:20.539 [SPEAKER_02]: And there are countless examples of features we built that didn't end up making it to production and I would say that cost us a lot of time and it would have been good to 20:26.394 --> 20:27.315 [SPEAKER_03]: So let's move forward then. 20:28.075 --> 20:30.937 [SPEAKER_03]: I'm curious about what the future is for strawberry. 20:30.977 --> 20:42.105 [SPEAKER_03]: We talked a little bit offline about some of the things you're moving towards and and the industry is moving very fast with AI and agentic AI and you and the team are poised right in the middle of it. 20:42.345 --> 20:45.467 [SPEAKER_03]: Come up with what the future looks like for the product the company and the team. 20:47.170 --> 21:02.218 [SPEAKER_02]: I think we are going to have this new internet where trillions of agents are communicating with each other, making transactions, making decisions internally, inside organizations, but also between organizations, between consumers and companies, and this 21:02.838 --> 21:31.985 [SPEAKER_02]: new internet is going to have such amazing upside if we do it right but there are of course a lot of things that a lot of risks and there are a lot of things that we need to figure out what frameworks do we try to implement to make this whole thing observable so we can know does an agent have access to do this or not and what what new protocols need to develop with this agentic internet to run smooth we're super excited about building a browser today 21:32.745 --> 21:46.393 [SPEAKER_02]: Since we can start bridging the gap between agents and the web by allowing them to use your existing websites, but going forward, we want to be the ones that come up with these protocols and introduce them to the world. 21:46.693 --> 21:52.617 [SPEAKER_02]: And when we start to see less websites being used, which I'm absolutely certain is going to be the case. 21:53.157 --> 21:57.319 [SPEAKER_02]: And then you might ask me, Charles, why are you building a browser if they're not going to be on a website? 21:57.479 --> 22:01.322 [SPEAKER_02]: I still think there's going to be a huge need for rich content. 22:01.862 --> 22:20.164 [SPEAKER_02]: So we're gonna have generated videos, generated games, pitch-tML, branded web UI, or sorry, MCP UI stuff, like all kinds of rich content that needs to have a human spectator, and that's still needs to be rendered under the hood by a browser. 22:21.025 --> 22:37.799 [SPEAKER_02]: So even though internet is going more towards this headless direction, where you have new protocols and agents communicating, I still think there's going to be a huge need for a browser like product for users to view all kinds of rich content. 22:39.060 --> 22:43.146 [SPEAKER_02]: In many ways, I like the similarity between when computers first came out, right? 22:43.306 --> 22:49.694 [SPEAKER_02]: And you had some people, of course most people in the beginning who had personal computers weren't connected to the internet, right? 22:49.754 --> 22:55.081 [SPEAKER_02]: You had your local files and maybe some games and you could code some things, right? 22:55.802 --> 23:02.388 [SPEAKER_02]: I think computers really started to become valuable when they could connect to each other over the internet. 23:02.488 --> 23:08.494 [SPEAKER_02]: Because then you had emails, you had websites, you had all of these services that became possible online shopping. 23:09.034 --> 23:11.516 [SPEAKER_02]: And I think it's the same with agents right now. 23:12.117 --> 23:14.319 [SPEAKER_02]: We have seen coding agents. 23:14.859 --> 23:22.327 [SPEAKER_02]: that use your local files, get stuff done, and I would say coding agents is almost solved, right? 23:22.607 --> 23:37.062 [SPEAKER_02]: It's almost a solved problem being able to code with an LLAM, but it's local, and we still have not seen the full potential of what happens when agents start talking to each other, 23:37.482 --> 23:54.097 [SPEAKER_02]: at a massive scale, and that's what we're really excited to explore with of course strawberry browser but then in the future a whole suite of products that is centered around this concept of the companions that work for your team and communicate with other agents on the internet. 23:56.058 --> 23:59.782 [SPEAKER_03]: Let's switch to you, Charles, who influences the way that you work. 24:00.062 --> 24:02.244 [SPEAKER_03]: Name a person or many persons or something. 24:02.524 --> 24:03.345 [SPEAKER_03]: You look up to him, why? 24:05.418 --> 24:10.979 [SPEAKER_02]: I have a lot of great founders here in Stockholm that I take a lot of inspiration from. 24:11.500 --> 24:25.103 [SPEAKER_02]: So I'm pretty good friends with the founder of Loveable Anton and he's much more busy this days than he was in the beginning for sure, but we would always, and we still do find times to go for walks, talk on the phone. 24:26.043 --> 24:31.065 [SPEAKER_02]: And I just think he's a brilliant founder and a very visionary person. 24:31.586 --> 24:38.128 [SPEAKER_02]: And for me, he's one of the people who made me realize that there's no limits to what you can do. 24:38.789 --> 24:41.750 [SPEAKER_02]: And you think you have all these limits, right? 24:42.170 --> 24:45.392 [SPEAKER_02]: I surely I can't climb that high or aim that high. 24:46.092 --> 24:50.654 [SPEAKER_02]: And then really you just need a person like Anton to say no, actually, you can do it. 24:50.734 --> 24:53.035 [SPEAKER_02]: And look, I've even done it myself, right? 24:55.405 --> 25:01.847 [SPEAKER_03]: Okay Charles, I got one more question for you, so you get on a plane and you're sitting next to a young entrepreneur who's built the next big thing. 25:02.367 --> 25:03.347 [SPEAKER_03]: They're jazzed about it. 25:03.367 --> 25:06.508 [SPEAKER_03]: They can't wait to show it off to the world and can't wait show it off to you right there on the plane. 25:06.988 --> 25:09.628 [SPEAKER_03]: What advice do you give that person having gone down this road a bit? 25:10.609 --> 25:19.891 [SPEAKER_02]: I would say probably one of the most important things that young founder needs to know is that they just cannot. 25:20.751 --> 25:43.738 [SPEAKER_02]: give up ever if they truly believe in what they're building then they just need to keep going because there have been so many times where we have been very close to pivoting completely or just giving up altogether and it's those days where you wake up and you don't even feel like getting out of bed that's how 25:44.578 --> 26:11.324 [SPEAKER_02]: depressed you are because it's so hard to find a company and everyone around you just sees how you're spending countless hours building this thing and they think you're stupid for trying to go up against the biggest tech companies in the world or whatever reasons right and for me at least in those times when people have come to me and told me this advice you just need to keep 26:13.464 --> 26:26.075 [SPEAKER_02]: It's been like a video game where you have three hearts and it's time someone gives you this inspiration you get one heart that can be consumed by a bad day and can give you the energy to keep going. 26:26.596 --> 26:33.182 [SPEAKER_02]: So that that's what I would tell a young founder if you just believe that you're working towards the most 26:33.862 --> 26:43.766 [SPEAKER_02]: ambitious thing that gives you the most amount of energy and inspiration, then nothing will be able to stop you, you just have to keep going. 26:44.386 --> 26:59.913 [SPEAKER_02]: Even if it means getting up and just working for a couple of hours and then going back and resting or taking some time off to think about things, each time we felt everything was against us and there was no point in continuing, we kept going 27:02.794 --> 27:12.821 [SPEAKER_02]: was the ultimate prize and each time the universe throws you a bone if you just keep going so that's that's what I would tell that founder. 27:13.682 --> 27:14.883 [SPEAKER_03]: I think that's excellent advice. 27:15.063 --> 27:16.084 [SPEAKER_03]: I couldn't agree more with that. 27:16.404 --> 27:21.448 [SPEAKER_03]: Charles, thank you for being on the show today and thank you for telling the creation story of strawberry browser. 27:22.148 --> 27:22.969 [SPEAKER_03]: Thank you for having me. 27:25.150 --> 27:27.592 [SPEAKER_03]: And this concludes another chapter of Coat Story. 27:37.520 --> 27:40.264 [SPEAKER_03]: code story is hosted and produced by no a lab part. 27:40.744 --> 27:45.170 [SPEAKER_03]: Be sure to subscribe on Apple podcast, Spotify, or the podcasting app at your choice. 27:46.131 --> 27:50.637 [SPEAKER_03]: And when you get a chance, leave us a review, both things help us out tremendously. 27:53.541 --> 27:54.602 [SPEAKER_03]: And thanks again for listening.