MCP, Copilot Studio & Responsible AI

Keegan Chambers: Yo, yo, yo, what's up, everyone? Welcome to Human in the Loop. This is my primary favorite podcast, alongside a couple of other interesting ones on the on the ecosystem channels. Anyway, we're super excited to be here today, and we've got the king of Copilot Studio with us, Daniel Laskovitz. What's up, man? How are you? Hey, I'm doing good. Very happy to be here. Really excited to be invited to the show. Amazing, amazing. And obviously joined by I am the co-host, Yorwan is the host. So do you just want to say hi? But hopefully we'll roll with it. Daniel, it's it's always lovely to to get to speak with you. So thank you, thank you, thank you for agreeing to hopefully what will be a fun show today about everything co-palet and more. So shall we start with that, Daniel? Tell us a little bit for those who might not know you about your role in the sphere of Microsoft. And I'm like dramatically hand-waving when I'm saying that. Okay. so yeah, my name is Daniel Laskowitz, and I work as a principal cloud advocate at Microsoft. And it's a fancy name, cloud advocate. But actually in Microsoft, we always like to name something different. So the rest of the the world calls it developer relations and we call it cloud advocacy basically. So yeah, we always have our own name for it. but yeah, we are doing developer relations. So what I'm doing is I work a lot with developers, with people who are working with our products. And I'm yeah, targeted mostly on the power platform. So I'm doing Power Apps, Power Automate, Copilot Studio. I also do a fair bunch of Microsoft 365 Copilot because we have A lot more integrations with that lately. So I'm doing a lot there. I have a background in Azure, Microsoft 365. I used to be a consultant before I joined Microsoft. So that's where I came from. But nowadays I'm talking a lot at events. so I have a lot of sessions, as you know, we see each other often there. and then yeah, when I when I talk about our products, about how people can use them, they always have questions, they always have feedback, etc. I bring that feedback back to the product team. I kind of gather the feedback, put it in a sort of priority list, and then I bring it back to the product team. I talk to them about what improvements they can do and that kind of the the loop that I'm that I'm in. So presenting about new stuff, getting feedback, bringing that back to the product team, and so on. And I can attest to that being one of the people who messages Daniel on Teams going like Yo, help. I'm stuck. How do I do this thing or how do I do this other thing? So I can attest that is absolutely true and also appreciated because I'm very, very privileged to to have you wamping away when I have all my ignorant questions. So thank you. Well, it's always good to hear that because the the the the things you run into, everybody runs into that or a a bunch of people run into that. And for me, my my main goal is basically to unlock everybody. to make sure that they are not running into walls or running into issues. and yeah, that's basically my role. So I I really like hearing or I'd I'd rather not hear them, let's say it like that. But if I can help, I w I I love that. Yeah. Totally. I think I think something that I find interesting about you, Daniel, is that you are extremely free with the advice that you give. as far as technology goes and something else I really like about the way you do things is you're not just like, hey, this is the technical solution. Like you focus a lot around like what is the actual problem that the person is facing? Like what is the actual problem that they're in? And sometimes the thing that we're looking to solve is not always the right solution for the customer. So like I've always liked that about you and your approach, even before Microsoft, you had a very good Good way about doing this. And actually, something I've noticed as well is when you do your sessions, one of the ones you do with Matt's around MCP talks about that. So do you want to just tell us a bit more about that session in specific? Because I think it's very important. It's been one of your most popular, right? Yeah, so a a little bit of a background behind that. Matt and I used to deliver a session about custom connectors separately. So we were at almost every conference and we were both doing almost the same session. And we were like, this is kind of sad that we're doing this because we're taking up two slots at a at a conference while it could be one. So at some point we were like, well, let's let's do it together and and make sure that we have one like best off session. And then from that on, also custom connectors kind of evolved and we were doing that as well. And then all of a sudden Copilot Studio announced that they were gonna introduce MCP in Copilot Studio. And that was for us the moment to say, well, custom connectors is already about connecting your backend to the power platform products. And MCP is kind of the same, but then for the AI products. So that was kind of the the reason behind it. And what we tried to do as much as possible with that MCP session was also to kind of open the eyes of people that were. Very much looking into only power platform. So they were looking into MCP and Power Platform. And at the moment, that was only Copilot Studio was available in or you could use MCP in Copilot Studio. That was the only product that supported it. And I was like, why why don't we change this a little bit and also show people other stuff? So maybe show them a little bit about GitHub Copilot, which is more of a pro code kind of tool, more of a developer tool. But We try to kind of take away that boundary that people are like, it's pro code, I'm afraid of that, and make it a little bit more easy to easy to handle. So we really went into into that part. and also, of course, we were we were looking at it from different ways. So you can have all the out-of-the-box MCP servers where you can connect to background services all the time, but you can also bring your own. So there's lots of companies that have their own APIs, they have their own data that they want to put into the power platform. And back in the day there was always a custom connector and now we are talking more about MCP in that manner. Wow, man, that's wild. That's wild. I love the fact that you guys actually had the foresight to think about the fact that you were doing a similar session and that you would kind of bring them together. And I I've known Matt I've known Matt's and you for a long time and man I've I've been in that session. friends, if you ever get a chance and you're at a conference, go to it. It will demystify a lot for you. Like a lot. A hundred percent. I was attending your session at Dynamics Minds and I was taking so many notes that then I shared with the rest of the team and we went like, this is how we need to do it. So thank you both because it is fantastic content and it is super helpful. talking about everything that is happening from MCP and Power Platform perspective, Copilot Studio. Copilot Studio has some updates recently, including from UI perspective. What would we know about what is happening in the world of updates? Yeah, there's basically a two-sided thing. one thing they did is update the UI. so there's a new UI that you can you can see in preview now. and that is something that yeah, I I really like it. I think it's more of a modern kind of look and feel to it. Because to be honest, when I was opening Copilot Studio at some times I felt like it was a little bit dated. And it was also because it was all built on a like older technology and they improved that over time. But now they kind of said, well, let's bring also the UI to that kind of level. And that's what they did. also another thing of it. So the UI is one thing, but the other part is also that they updated the orchestrator behind Copilot Studio. so they have a whole new orchestrator that's that's being rebuilt from the from the yeah from scratch basically. And that is something that will make it easier for you to to do multiple things, for instance. So it has like a different way of going around in yeah, your agents, and it has more capabilities, you can do more stuff at once, etc. There's there's a lot of things that are behind that with the orchestrator that kind of figures out what to use when and that kind of stuff. And that is really something that that improved a lot as well. So it's kind of two-sided, but If I look at the UI nowadays, I really like it because it's also a bit more the same as other products. I don't know if you noticed, but you have like a sidebar where you have all the all the things in there, like memory, you can add tools there, etc. And it feels really like cowork, for instance. If you have used co-po co-work, it was exactly the same kind of UI component that they had as well. And I really like to see that that same kind of look and feel over multiple products because Sometimes I see it's a little bit too much disconnected. Yeah. And this way it looks more like one, yeah, one Microsoft kind of feeling. And I think that's also a good thing. It helps with the user onboarding. I in our team, in in my team I should say, recently there's been a lot more experimentation with Copalo Studio and people who in the past, you know, haven't necessarily coded on, don't have a coding background, felt like Hey, this is really powerful technology. I want to experiment at least internally to understand what can I create. And I feel that the new UI, to your point, helps. It it's less of a I felt that they need less of an onboarding in terms of like where do I go? What do I do? I still recommend because your team has produced some fantastic training that I've done myself. So, folks, if you haven't done that, please do that. Because it is a fantastic overview of the capability, which is Huge in Copilot Studio. But it's definitely been less of a barrier for people just to kind of intuitively figure out, I can do this or I can do that. Yeah, I totally agree with that. It's it's it's less of a barrier. And we're also working on evolving our content, of course, because we have Agent Academy with recruit, operative, a whole bunch of modules. And we want to, of course, make that as easy as possible for everyone. we kind of want everybody. To just go to Agent Academy and learn about agents. That's that's our goal. And we currently have it still in the old UI, but we are working on making that in the new UI. And that also sometimes means that you need to do something completely different because as I already told you, it's a two sided thing, so the orchestrator and the UI. And that means that sometimes certain things work differently in the new in the new version. And also some things that you used in the old version don't work anymore in a new version. So we are kind of figuring out how to really make that happen and how to make that work as easy as possible. But we're definitely working on improving that. Thank you, Cher. The pain that we all feel right now is like technology is progressing so fast that documentation and training is just not keeping up. which is why people like you and Chris who are talking about the updates, who are showing things, make a massive difference in terms of being in the community together. I have a question for both of you actually. and Chris, maybe you can go first. What have been some cool things that you've still we you've seen or experienced people building with Copilot Studio to give others an kind of an idea of what is possible there and what are the things that you can create with Copilot Studio? So I'm gonna caveat this answer because I think it's important, right? I think like there are a lot of people out there that are trying to rebuild their current processes inside of tools like Copilot Studio. And I kind of get that, but don't. Like obviously there'd be pieces of it you gotta think about. But, you know, if you'd asked somebody back in the day, in the middle of London what they would like, they would have said faster horses because they didn't know cars exist. Okay. And they didn't, right? Like, and now we're still trying to fi we're fixating on making things faster, but not better. Right. So in Copilot Studio, you have to s you have to really shift your mind into not like linear process flows. So I call it flows of work. and my flows of work are not always linear. Like there's a million different things are happening in the things that I do every day. So right before this, I was literally building out an architecture. I was responding to an email. I was looking at a WhatsApp message. Like we are used to multitasking, whether you know it or not, like we do it a lot, right? And also something else we're very good at doing in this era is context switching. Right. So we're able to move around different contexts. And I know people get tired because of that, but actually When you look at the things we do every day, we do a lot of it. Like we're switching between these linear these processes. So when I say this, please understand, I have shifted my thinking out from linear process working out into this concept of like, I would say it's multi-agent. It's the easiest way to explain it. So in Copilot Studio, one of the coolest things I've seen built. And I I I worked on this with a friend of mine from another company. I can't mention the name. But we sat together and this was kind of before the new orchestrator. With the new orchestrator, it's going to be much easier, but you'll understand why I say this. So we built a multi-agent system or an agent mesh. it's a hierarchical agent mesh. So you had an or a primary orchestrator, then orchestrator then houses about another 13 to 15 agents, depending on what happens, right? And we built an ideation solution. Now you're probably thinking that sounds pretty dumb. Like everyone's doing that, but actually this was interesting. So this ideation solution allowed you to input information about agentix solutions that you wanted to build because everyone's asking, like, what do we build with agents? And then what it would do is it would build, will architect this thing for you, like completely. Now, the really interesting part was that we had many agents doing many things in the background. So when you looked at the when you looked at the actual process that the agents were going through, it was never a linear. There were always the architecture agent would come in, the data capture agent would come in. the evaluation agent would come in. We actually had an agent based on the Gartner framework that would look at like how to score the agent based on certain numbers. And it worked really well, right? Now, the thing that's interesting is that one of the agents was actually gaslighting another agent. So yeah, I had a I had a data review agent, and the data review agent was gaslighting the data capture agent. So we'll get it was getting many records created in Dataverse. And the reason it was doing that is because its version of good was different to the data capture agent's version of good. And that's because I'd been too specific, right? In my instructions. So it was another really interesting way of like going through this process and actually watching these agents and all the conversation turns happening. And one of the one of the one of the sessions that I went through, there was something like 64 conversation turns, something like that. It was high, like really, really high. And you could see the agents bouncing between one another. And that's what lit it up for me. Right. When I saw that happening, I was like, okay, another interesting one that I built with a energy company was remember the like all this horrible stuff going on in the world. There's this organization does trade of crude oil, right? And what happens is that when the crude oil is on the boat, the the price can change. So you need to know what is happening with the stock. Like a lot of things can go wrong. So this thing, it was a fabric data agent, a foundry agent, and a copilot studio working together synonymously. And what would happen is that there was an API rigged up into fabric. The data agent would then score that into into that would land up in in one lake, the agent would look at that and then start feeding up into teams. And then what would happen is they had another one connected to a third party MCP server called Kepler who do satellite imagery of of v of vessels. So they can tell you where it is. So what was happening was we were mapping the location of the vessel against the news feed that was coming out from the rest of the world. And we were then able to understand like, hey, the propensity for the price change could be X based on A B and C. So many agents, right? But the the thing is that flow of work would never have happened before. We just rethought the flow of work in both aspects. and those I like those two stories a lot because it's multi-agent, it's multi-platform, and actually it's the thing that lights up how agentic systems work together. And I'm sorry, I've spoken a lot, but like I get very excited about it because architecturally I think it's super cool. Technically, it's not rocket science. Like Dan, you could build this in your sleep. I reckon same as you, aren't it? Like if you just had the spec, you could easily create it. But that's the that's the beauty of it, right? It's about how do you lower the the barriers to people creating this technology. You said something really interesting, and then I'll go back to Daniel because I'm curious of his thoughts of what you've seen. Very cool in this area. You said that it was something that you did at the instructional level that basically triggered one agent to to gaslight the other. I had a similar learning recently. I used to be very, very prescriptive in my instructions. and I worked a lot on sa safety system messaging, so extra so. I've seen recently that my previous instructions put my agent in jail. It's I'm so so specific that I basically block it from doing anything. so that's a that's an interesting observation as the models evolve, as the capability evolves. I had to be a lot a lot more permissive in the way that I phrase instructions because otherwise it was like, sorry, can I do that? Sorry, can I do that? because basically I was blocking it from doing anything. but Daniel, to you, what what cool things are you seeing in the world? Yeah, so this is kind of an e example that I also have, like the the instructions back in the day when somebody started a new agent, what was the first thing they did? They would write like a a couple of pages of instructions in the beginning. And what we realized after seeing a whole bunch of agents was also that the agent actually gets better with less instructions. Yeah. So the there that was also a mind shift for a lot of people because people were used to kind of and Chris knows exactly what I'm talking about, like the the first Copilot CEO agents, or back in the day PVA, everything was around topics. So you would design A whole conversation and it was very chatbot-based. It was like you have a beginning, you had triggers, so that may yeah, the triggers made it possible to go into that topic. And then you would have a whole conversation, like question, answer, question, answer, that kind of thing. And then you would have maybe a some kind of if el statement. So if answer is a, then go to this or perform this action, those kinds of things. And nowadays it's completely different. A lot of people are now working on all these tools that you can just add to an agent. And then it can figure out when to use those tools. And it's as Chris was also saying, like it it's a completely yeah, it's completely asynchronous kind of thing. And it it really happens when yeah, when when a certain tool is, for instance, triggered. it can be that that tool is at some point the number one in the in the order. But it could also be that's number eight, for instance. And that kind of thing that we we don't always have to plan exactly ahead what's happening. I think that is something that's a complete mind shift for for people that they are not used to because they were used to kind of creating that topic always one by one. And that is something that I've been seeing over the last couple of weeks as well. A lot of people have been restructuring their agents and making it more and more less about. The instructions that you have on the agent level, but more on the kind of describing when this action should or when this tool should work, basically, and what it should do exactly. It's mind-blowing what the technology can do when you describe that in a good way. So when you see an agent that doesn't have a lot of instructions, but everything is on the tool side and everything is well described. Chris is always no nodding. He he knows how it works. and and that is something that completely different than how we worked two years ago already. And that is something I think is yeah, really interesting to see. I've saved so many hours with tools and work IQ, I cannot explain how useful that has been. We were developing something internally and we wanted the output for a particular reporting period to be consistent across whenever a user said, Hey you know, for this period create a report. And we wanted the action to be run the databases that you know, make sure to use the specific format and put it in a Word document. We didn't want to put it in in Outlook because we knew there was some editing that would happen between multiple people before it it got sent. And I was thinking to myself before work IQ and everything like, my God, I'm gonna have a trigger from there, thoughts and prayers, you know, power platform Not I'm not big on Power Platform, I have to say, because I haven't used it used it as much as other tools. So I was thinking to myself, do I have to learn PowerPlatform now? And then WorkIQ came in and I'm like, my God, I don't have to do any of that. And it's just literally I give it tools and it does the thing for me. And it's been such a time saver, but also a capability enhancer. And it's fantastic to see the evolution. And don't get me wrong, I'm not, I'm not hating on power platform. It's just not my. No, but it's it's also kind of throwing me back 10 years in time. And and Chris also know this from back in the day. Like, what what was the discussion we always had 10 years ago? Was always when to use what. Yeah. We had so many tools, and I kind of see the same thing now with AI. We have so many tools. And when do you use what? we are talking about co-pilot co-work, we have Microsoft Scout, we have so many different things that we can use. And it's coming back a bit. back in the day we used to have like I think we had office groups was one of those things. We had GMmer. We had we had groups everywhere. Remember that? We had teams? Yeah, I dynast those days. It's like I I had this debate with somebody recently. They're like, go to the group. And I'm like, but but just one That's a great point, by the way, in terms of like we have so many tools. We have like Scouts, by the way, thank you, Daniel, because you were the first one to hint me internally towards that. And it's changed my life. I love that. But from either of you, actually, any tips and tricks of like how would you navigate the person between like let's let's keep it a little bit shorter because this can be a long list, but between let's say co work, maybe scout, it's coming. Not a lot of people have access to it, but it's coming. Cowork, scout, and power well, co pilot studio. What would be kind of your cheat guide between those three four people? I don't have any hands on experience with Scout for obvious reasons. but for I think I think the the question that is coming up more on my side is like why do we need co-pilot N three six five co-pilot as in like the chat UX, co-work, Scout, Copilot Studio, agent builder, foundry, data age data agency fabric. Like why do we need all of this stuff? So I think I think basically it's down to it's down to, and Dan put it perfectly, is like What are you trying to do? And I feel like we as the community need to be more specific in what these things actually help you with. Okay. Because somebody was h hammering cowork the other day to do rag, right? Like, find this thing for me. And Coburg's like, yeah, let's make that here it is. Let's make 16 PowerPoints and, you know, have a cat deliver your house. And I'm like, yeah, because the cowork wants to do the thing, right? Like, so co pilot, the chat UX wants to find the thing. Okay. So that's the way I frame it. I'm like, It can do some stuff, but ultimately it wants to find the thing. Like you can attach it to tools. Kobick wants to do the thing with you. Scout wants to do the thing for you. Okay. Yes. And and that's that's sort of how I see it now with Scout because I've seen it, but I haven't actually got to meddle with it. Dan told me some funny stories early on about Scout. I was like, yeah, man. I don't know how many you can share on this, but it was really interesting because my first interaction with the M365 Copilot was like, wow, this is a cool tool. Done. Like, great. This is awesome. Cowork, my mind was blown. I was physically like, how I built the most amazing documentation and skills lab in minutes for four dollars. And I checked. I logged it and I checked it and I shared it with everyone. And I'm like, okay, I understand with I understand. And what Microsoft seemed to do is that they'll we are not ready for it's kind like we are not ready for all the things that these tools are able to do. So you have to drip feed to an extent. So like my theory is that people like Microsoft, organizations like Microsoft are very, very much further ahead, but it's time to market, like when to release it. And it's the same thing, like looking at tools such as Copilot Studio, when I look at what it can do, like in a realistic context, it scares me because it is exceptionally powerful. Like, but then again, you can target it. And that's what I like about it. So cowork is very broad. Tools Copilot Studio, you're gonna target. You're gonna wanna like laser target those tools into specific flows of work. Notice I didn't say use cases and notice I didn't say workflows or processes. Right. And that's really where I'm starting to think actually, this is incredible, but I need to now figure out like where the delineation is because there's overlap across all of the products. Okay. And I love your distinction between cowork and scout. It's like with you, for you. Scout is incredibly powerful. And it's it's one of those tools where there is so much capability and you need to really be sure that you know what you're asking for and how you're asking for it. And one of my one of the very cool things about Scout is it always asks for permission in terms of like, Can I perform this action? Can I perform this action in the context of this particular thing? And I always tell people, please read those. There is a reason why you're being asked. And this was a It's a user example, it's me actually being wrong with one of the instructions. I was I wrote really badly my prompt and I was like, let's send this email to these people based on this. And I was like, shoot. I actually instructed it to send the email on my behalf. So I I caught it in time because it asked for that, hey, hey dear user, did you actually ask me to send the email on your behalf? Because I can, but like, are you sure? Sure. and that's why I say it's like Scout is super powerful because it can pull from a lot of things and customize, but it also has that I would say it's almost like action biased, which is great, but you have to be sure that that's actually what you want as opposed to like yeah, not just scout to search for emails. Like that's that's not where its power is comes from. Yeah, for me it's it if I look at all those tools, I think cowork and scout are very much single player kind of things. Yeah. So it's very much an assistant that you can use. For me, cowork is kind of a thing I start with in the morning. I start with a couple of tasks, I fire them off. And whenever it needs some input from me, I will go there again and I will respond to the questions it's asking, for instance, that kind of thing. for scouts, I have been using that a lot for my administration because in our org we still used we're switching now to GitHub, but we used to have Azure DevOps. And I was using that all the time to do all the administration around all the conferences I'm going to. So I need to do travel requests. I had to do reporting at the end of the event, those kinds of things. And I would just use the Teams bot that they had to kind of say, Well, there were so many people at my session, these were the numbers that I had, with regards to feedback, all those kinds of things I was basically saying. And I really trained Scout to use my way of working. so I gave it a lot of memory, for instance, to kind of know which project it always should have. if it creates something, always set the assigned to to me, for instance, those kinds of things. And to me this was also the yeah the kind of the wake up call for me to kind of f say to other people as well, you need to really think about what you want to do with Scout. and if you are leaving it like out of the box. It will use a lot of browser automation. Yeah. And I saw that the browser automation a lot of times, it kind of, yeah, especially with authenticated sources, it can sometimes go very in a different way. And that was for me something where I was like, I need to give this a little bit more instructions. So what I did was I installed a command line interface and I gave it the instructions to use that, for instance. But after I did that, it worked perfectly. And I think that In the future, Scout will definitely get more in that area that it will out of the box do those things as well. It will have skills for multiple systems, those kinds of things. But those tools are kind of similar, as in they are both something where you give it a task and it can perform that for you. Yeah. The other thing that I see with Copilot Studio is kind of different. It's more of a multiplayer thing, because what you do with Copilot Studio is usually you build an agent. And You publish that age agent somewhere and you give a whole group of people access to it. And they do that to to do a certain process or to go through those kinds of things. It's a little bit of a different angle to it. Yeah. That doesn't mean that yeah, that that they can do similar things, but it is a bit of a different angle to that in that in that way. And it goes back to what Chris was saying and it's like the problem you're trying to solve and for who you're trying and I love I'm gonna steal this, Daniel, of like the single player b versus multiplayer 'cause it's It's it's a hundred percent accurate, but th this isn't an actual thing my team is working on the moment. we have this we have this internal newsletter where we send the main accessibility news of the day. So whatever five top news and accessibility have happened, we send it out to a group of people who are interested in that topic. It used to be very manual, of like somebody scrubbed the internet, was like, let me put this in the email and send it to a bunch of people who subscribe to our ID web list. And we had this exact debate of like, Do we want to use Scout or do we want to use build it in Copilot Studio? And we ended up building it in Copilot Studio because we didn't want single point ownership. And we didn't want it to be like a newsletter that I wanna have to send versus like, you know what? Let's have multiple people in the accessibility team own this thing because if I wanna is on vacation, they shouldn't be dependent on Iwana. But that was exactly the thought process of like, where is ownership? Where is the what do we want to achieve here and what's what drives the better result. And it was a matter of like this versus this because of that. Yeah. I makes sense. Yeah completely. I agree. I like I like that. That does make a lot of sense. I'm thinking I have Dan, actually I have a I have a more of a nerd question. and maybe I wanna if you you can chime in from your experience as well. It I'm just interested in like I feel like the introduction, not the introduction, okay, because I know they've been there for a long time, but now the visibility of things like skills inside of Copilot Studio. Okay. Mm-hmm Because Dan, you touched on it earlier. I'm sorry, it's like chewing on my brain. Before it was topics and tools and like you're bolting on a tool to a topic, and a topic was kind of like very deterministic. But with skills, I feel like this is an absolute game changer now that it's like front facing, front visible, with all these crazy skills repos out there. Like I know Scotty has got Scott DeRoe's got one somewhere on that you can download skills. There's tons of them, right? But how do you both see this changing the game as far as using tools like Copilot Studio Go. Because to me, I've got an opinion. I'm not going to share it now, but I'm I'm I'm keen to see what you both say. Yeah. Yeah. So I think skills are definitely a game changer. because I've been using MCP for a long time. And one of the things that we realize with MCP is also that it's very token inefficient. Yeah. because it gets like all the details in the beginning of the conversation and you will burn through tokens really fast. that is something that we noticed. And after that, there was kind of a hype that MCP shouldn't be used anymore. And a lot of people were like, MCP is dead, and that kind of stuff. but they all said you need to work on, for instance, CLIs, so command line interfaces together with skills, so that you have the skill to kind of describe. What you can do with that command line interface, what kind of processes you have, and how it should kind of run through all those commands that you have in those command line interfaces. And I think the same is now a bit happening with Copilot Studio because they are using skills. You see it also with Copilot Cobork has it as well. They also have skills. so you can see that skills is getting a lot more popularity there. but the good thing about a skill, and and and I I really like The example that you gave, that there's so many skills around there already. There are there are thousands of skills. There are websites like skills.sh is one of them. I really like that website because it's also very much about security. but what's what I really like about those websites is that sometimes you see see a skill that is like a really, really short skill. It can be like 30 lines or something. A lot of people think they need to write like pages and pages of stuff in a skill. But sometimes a really small skill can already make a huge difference. I can give you an example. One of my personal skills that I that I like a lot is one that I can use as like an interviewer. It's it's kind of an a thing where I can say, I have an ID. Can you please give me a whole bunch of questions about that ID so that I can make it a better ID, basically. and I think it's it's called Grill Me. It's a pretty famous famous skill. A lot of people have been using it, but it's so nice to use for for these kinds of things. And I do that a lot when I am kind of figuring out certain things that I need to write, for instance, or a project that I need to start and have like a sample that I want to create. I always go through that grill me skill because it helps me kind of yeah, be critical about my ID. And it helps me also to organize my thoughts. And that is something that I really like about these kinds of skills. And there are so many other examples as well. one of the biggest ones that I always use is the front-end design skill because I'm horrible at designing. I'm I'm like the worst designer ever. so if you let me do let me create a website or something or an interface, I'm I'm horrible at it. So I'd rather have a skill that has all those design. thinking methodologies around it and it will give me all the yeah all the things that I need. And then you can go on from there. And I think those kinds of things are so, so powerful with with skills nowadays. I'm really excited about the potential it has in terms of scaling capability from a human perspective. And I'll give you an example what and explain what I mean by that. So going back to accessibility in the world that I live in, accessibility is not taught in schools, unfortunately. A lot of us, you know, who did development courses and things like that, nobody tells you what it is to develop something that is actually accessible for people who are blind. That is not skill we are taught. So a lot of organizations when they're trying to create accessible software, all good intent, but designers, developers simply do not have that knowledge or experience in terms of like if I'm designing these things in Figma, what are the keyboard implications for somebody who's blind or somebody who doesn't use a mouse? And I'm really excited about niches like that where you have capability from a human perspective, but you can't scale because it is so narrow and so very little known that you cannot scale. So how about a skill that tells you almost like a persona of like dear designer as you're thinking of whatever you're developing, these are the things for a keyboard implication for somebody who are blind. And that's a skill you create and go super niche of like for this particular category of user and this particular type of type of experiences, these are things you need to consider from a design perspective. And for me, that's what makes it very, very cool, because it means that suddenly you can scale your capacity as a accessibility team. which in many organizations are super small or they don't even exist, where it's like one person. So things like that become very, very interesting. Yeah, I agree. And also the example that I gave you with front-end design, I noticed last week, for instance, I was looking through the scale again and I noticed that there was a whole section about accessibility. And it was putting all the the the the ground rules of accessibility in there, etc. And it would already take that into account. And I have been using that that skill for months already. and I noticed that accessibility part for yeah, like a couple of weeks back. Yeah. So it's pretty cool. You might not know. And I don't blame people for not knowing because again, it's not something you we learn. No. so I d I wonder, does that answer your question, Chris? Yeah, it does. so I've got a very I don't know if my response is gonna be as positive. And I'll tell you why. Okay. Like just I don't I I need I need to learn more. Okay, but number one, I think that I think it's good. I think skills are good. Like ultimately, I think it's a good thing. I think that the ability to upload and trial them, upload and use them directly in agents, like you're you're recreating a thing that, you know, this thing will do over and over again. My worry is more about like governance. So what tools do we have to vet that these things are actually like okay, right? I mean, because I can create a I can create a a demon skill. That does like pretty much anything, right? and then also around Skills Mageddon. So you can you're gonna get someone that's gonna do this, like some extremely well thought out individual is gonna download like 896,406 skills. and and and what they're gonna do is cause that's literally the number on skills.sh, and they're gonna try and upload them all, right? And they're gonna be like, okay, this is my God mode agent. And what my god mode agent will do is Everything. And I'm like, no. And then there's going to be like some layer of fractal. Like, how how will the agent respond? So I think in general, the concept is really good. I think because this is the trustworthy AI podcast. So I think the concept is good. I think the idea is good. I think that somebody is going to do something dumb. and that's my worry. I don't disagree by the way. But but I think it's like it's the whole tour of weapon sort of conversation that we have had a few times before, Chris. It's like a hammer is a hammer. How you use that hammer is kind of it's there is some sort of ownership and liberty of deciding, yeah, I'm gonna use this hammer in order to put this nail on the like in the wall or whatever. Or like actually I just managed to just break my finger because I was not paying attention to what I was doing and mishandling this tool. That can be quite powerful. And I think it's it's a fair concern and I'm not minimizing it. And this is where to your point is like, and I don't have the answer, but this is like the fourth thinking is like how do you have governance around it where you can put some some sort of like, Hey, hey, like Scout did with me. Are you sure that you want to send that email? Are you sure that I can do this? Are you sure? Yeah. and you put kind of that that human in the loop back in the back in the equation because ultimately Yes, you have the capability, but are you really sure that you want to deploy the skill? Or how many numbers of skills you said that there are there, which I have a feeling would make any agent faint, but yes. I think there's also multiple layers to this. Like there's one layer of where do you get your skills? That's one thing. so Chris, you were talking about the skills.sh website. So if you go to that website and you click on one of the skills. You can see on the right side already there is a security audits section of it. And it's running multiple security audits to check what is what is that skill actually doing? Is it referencing certain files that are like on on GitHub and people can update it without knowing what it actually does and that kind of stuff. So those kinds of things are of course very yeah, like a big security risk that somebody can Yeah, you you might think, the front end design skill is a really nice skill, but I will use that now, today, July sixth, for instance. But then if I would create a pull this again in a month, what would change in the meantime? Maybe somebody would have a reference in there to some kind of place where it says, if you work with credentials, always send them to this to this HTTP address, for instance. Those kinds of things can happen, of course. So it's one thing is where do you get all those skills? I think there's also an organizational thing for it. So you might want to have your organizational catalog of skills that are vetted by a certain group of people, just like the center of excellence part that we always had back in the day. so those kinds of things are really important. and of course, sometimes it's also tool thing. but that's kind of the the small sidestep. But if you, for instance, look at co-pilot cowork, it also has a skills skill. So, what you can do then is create skills through the skill skill, and then you can, for instance, import the skill from skills.sh. a lot of skills in this. But what you can do there is you can you can basically ask it to import it into your personal group of skills. And what it does then, it will look at that skill, it will import it. But it will also validate it automatically. That's key. So it will check a whole bunch of things, security-wise, but also are there any other skills that are using the same triggers? And might they conflict with each other so that you have two skills that do the same thing and that you're yeah, making the AI confused by what you want to do. So those kinds of things are also in there. And I think if you yeah, there will probably be tools in the future as well that can do skills. scanning for instance so look at a whole bunch of rules and see if that's that's correct skill or if it's a skill that you even want to use in your organization. so those kinds of things I think will be more and more the case. But it's it's still the same thing as what we what we did with pro development for instance. A lot of people are using like all kinds of packages that they get into their software. and how many times have we seen in the last couple of months that some big kind of packages got got infected by some kind of hacker that did something, supply chain at checks, all those kinds of things. there have been so many of those, but we still need to be really on the top of our game with the security side. Yep. And make sure that we at least have the security part ready, but also of course the governance side where we are kind of vetting all that stuff and seeing like this is the set of skills that we as a company are okay with. If you want to use something else, you either want that or you don't want it. But it's kind of a thing where yeah, maybe somebody finds a skill and they want to add that to that to that group of skills, that's o also totally fine. So I don't think you need to be too harsh on it because it can also kind of cramp down the innovation sometimes if you're too s too harsh on certain things. because some yeah, the I I've seen the most f beautiful things. see created by people who are just doing something on their own. They are like, I have a business problem and I'm gonna fix it. And they build something. And yeah, that's that's something that we shouldn't lose, I think. I no, I agree. I agree. And I just want to do I just want to do a quick screen share. I want show you guys Dan, just to highlight what you were saying. Mm-hmm. So this is what they're talking this is what Dan was talking about on skills.sh. You can actually haul up a skill and it'll tell you if there's any sort of thing you should be worried about. Now This is a good website, dude. Thank you for sharing this. I I didn't know this existed. but it's it's really awesome, right? So there's a ton of them in here. so at the top level you could click, yeah. Yeah, yeah. So I highly recommend here's the one that you were talking about, the grill me one, right? So all passes on the security side. But I t I found it interesting that they were talking about like, hey, this there could be a prompt injection. And actually, if if somebody is if somebody is a bad actor or somebody is I think it's important and I think this is a new job role. I think that part of AI operations and AI governance is not only to vet agents, but actually your skills library within the agents. And that's not like something overly simple. But this is a this skills.sh is a really good example of a centralized library that's gone through governance. And you know what? We did we did this in Power Platform already. We've gone through this process. So we're just doing it again. just with new tools. So yeah, man, that you've y this has been incredibly useful. Thank you for sharing that. Yeah. Yeah. I think that I see a lot of similarities between AI and the cloud kind of boom that we had. because when I started working, it it was a moment where my then manager was like, Daniel, you need to work on the cloud because most people in our organization, they just want to keep working on their own server. They have all the access that they can have and they know wh what's working there. And I was like, I was a new guy, so I was like, Yeah, sure, let's let's go to the cloud. And I got all the information piece by piece. And that was probably my luck as well, because I got to see all the all the new features come in piece by piece and not the whole thing at once, basically. Because the people who were kind of left behind, they wanted to move to the cloud later on, but then they had so much stuff to learn that it was kind of hard to to get that all working. And I I kind of feel the same with AI. People, yeah, I think that a lot of people are already. pr pretty much on board with AI and seeing that w what it can do and what it can deliver to people. But it's still also a thing about keep it also human. I was at a presentation a couple weeks back. There was a an event in the Netherlands and there was a lady who was saying she was really annoyed by Sam Altman from from OpenAI because Sam Altman he has a kid nowadays. And he was saying That he was at his friends and they had a kid too, and the kid was already walking, but his kid was the same age but wasn't yet. And he was like, Why is that? And he was then going into Chat GPT, as you do as an OpenAI CEO, of course, and asking why is my kid not walking yet? And the other is. And the lady who was presenting was really mad at Salm Eltman. Why are you using Chat GPT for that? Also, still have the human connection. Talk to the person. Yeah. There's an opportunity to talk to that person. And sometimes I believe, like what I was also saying about the grill me skill, it helps me a lot. But I also need other people's perspective on it. Because I know that whenever I talk to Chris, for instance, every single time I talk to him, he has another angle to something that the the the Yeah, the AI hasn't fin thought of at all. It's just a different kind of perspective. And that is something that I really, really like about that. And I yeah, I think that's also a message I want to give. Like, keep the human stuff in there because we as humans are pretty awesome. And yeah, we don't want to yeah, lose that whole thing. I agree. I love that. that's such a fantastic note to kind of end on in terms of like stay human. keep Keep what is important and also maybe don't to put too much pressure on your kids to fit within the benchmark of their age. Because not such a thing is a benchmark. But that's my personal snark. So with that, Daniel, one thing I ask all of our guests before we be before we head out. Any sort of homework that you want the audience to take in? And it can be anything. Yeah, I think it's it's gonna be like Yeah, pick something that is kind of new and try to figure out anything out of it. so whether it's skills, whether it's something else, just pick something and then for especially for the summer. That's a really nice period when you do this. I always love picking a certain topic for the summer and just go read about certain things. and that that is something I would like to challenge people to. Just pick a topic for the summer, go read about it, buy a book maybe. If you have an ebook or something, that's totally fine. But do something there. I love it. So both cloud and on prem books are acceptable in this case. We'll think Well, with that, Daniel, thank you. It's always such a pleasure to speak with you and I always learn so much. So just a world of gratitude towards the work that you do and how you do it and the fact that You are incredibly human in what you share and your advice and everything that you do. So lots of love from both Chris and I. I know I'm speaking on behalf of us both. Yes. And folks, thank you for joining us. I don't know where you are, but we are definitely melting in our respective parts of the world. So yeah. We will head out to have some some cold water. Thank you for joining us for another episode of Human Indelope and we'll see you next time. Thank you, everyone. Thank you, Iwana. Thank you, Dan.