You’re not the user anymore | Layer’s Andrew Hamilton

0:05Welcome to Dev Interrupted. I'm your host, Andrew Ziegler. And I'm your host, Ben Lloyd-Piereson. Today we're talking about Google donating their agentic protocol A to A to the Linux Foundation, why everything wants to be your new IDE, the newest entrant to look who's AI native now, and an update to one of last week's biggest stories. Ben, you know the drill. You know what? Let's just keep it simple and start right at the top because I love the Linux Foundation and I want to hear what's going on over there. Yeah, so there was a recent announcement that Google donated the A to A protocol. That's the agent to agent protocol to the Linux Foundation.

0:44This was a formation actually of a project within the Linux Foundation, which is their typical approach to fostering an ecosystem of tools. And this is a project that they're working on together with AWS, Cisco, Google, Microsoft, Salesforce, all of the big players, really. While it'll be hosted on the Linux Foundation, it'll be seeded by Google's background for having created the tool. Now, the A to A protocol, if you're not familiar, it's an open standard for communication and collaboration between distinct AI agents. You can think of it as like a language that allows agents to speak to one another, commit information back and forth, and it will be another critical step in orchestrating the AI agent platform of the future.

1:26So I'm interested to see what this means for development for A to A, how the stewardship might evolve as a community that's hosted in the Linux Foundation. I think it opens up a lot of doors for innovation here. I really love the Linux Foundation, as I mentioned. I briefly worked there many years ago. I have a lot of friends over there. It's a great organization and truly just wonderful stewards for open source projects. And I think the best testament to this is the fact that they got Microsoft to get on the good side of open source software. I'm never going to forget when I first saw Microsoft show up to a Linux Foundation event with all of their flying pig, like swag, because nobody believed that this company that previously called open source of cancer would show up in that way.

2:08So yeah, Linux Foundation has done a lot of great work. And I think if you're out there like wondering, where is AI headed in the next five years or 10 years, maybe? Like, so looking a little further ahead than I think we typically are just because of how fast this technology is changing. But if you're looking ahead, I think this is where we're going to see the real productivity and the real fundamental changes happen from AI. So I think watch this project, watch what's going on here, because much like how HTTP created a lot of new opportunities via the internet, I think there's a similar potential here with A2A when it comes to agentic AI.

2:45So I think the potential impact here is pretty extraordinary. And the fact that it's now like a neutral organization within the Linux foundation paints a pretty bright picture for its future. So speaking of Google, I guess we have a few AI updates that we wanted to talk about. And the first one I think is Gemini CLI. So what can you tell us about that, Andrew? Yeah. So Gemini CLI just dropped in the last week. It's your open source AI agent from Google backed by Gemini. It's another entrant into the agentic coding tool space. These are tools that you can use in your CLI that allow you to interact directly with your tools.

3:22They are agent conversations that have tools and can operate on your system. You can use these types of tools in lots of different ways. Some folks use them in the IDE, like my daily driver's cursor. But lately, I've been actually trying out a lot of these coding agent tools in the CLI. I like to live there sometimes. I've had the luck with Cloud Code and GitHub Copilot, so I'm definitely going to be checking out Gemini. I also recommend Sourcegraph's AMP. All of them are really well-formed coding agents. Though I will say recently I've been using Goose probably the most in like one off sessions with local AI.

3:56And I also use an MCP server with it. But I've been practicing on Goose because I'm actually going on the great Goose off later today. This is an event that happens through Block, the company that created Goose. They have a YouTube vibe coding show where they bring folks together to try to build things on Goose on a YouTube live stream. So I'm going to there probably right after this, actually. And if you miss it, be sure to go check it out on YouTube and see how I did. Let me know what you think. And most importantly, if you haven't tried out these tools, it's worth experimenting with this one from Gemini.

4:27The Gemini CLI's free tier has a high usage limit, and it's the highest in the market, actually. So it's a great way to try out agentic coding for the first time. Yeah, you know, it's just yet another company building more AI features to bring feature parity to the competitors. But I like how this does tie to our first story a little bit in that everything is sort of going to get an AI agent attached to it. And they're all going to be communicating with each other at some point. And the CLI makes a lot of sense because it's where developers spend a lot of their time. And yeah, if you're catching this live, make sure you tune into Andrew live right after this or catch it on YouTube after the fact, because I'm sure it'll be a lot of fun.

5:07I think we also have an update about a story we covered last week. So what's that, Andrew? Oh, yeah. So we talked about how Meta was going after OpenAI's researchers. This was an alleged thing that happened that we talked about in the news last week of Mark Zuckerberg trying to take researchers from Sam Altman's company. And at this point, we might have to start a scoreboard because in the last week, we've learned about three such researchers that have taken that deal and have moved over to Meta in Zuckerberg's ever-growing quest to reinvent Meta's AI by trying to get as many experts as it can.

5:39And for how much these engineers and these researchers are getting paid, it's like big numbers, Ben. It reminds me of like big league sports in a lot of ways. Like at this point, I feel like we need like an ESPN channel on this. The signing bonuses make me think that all of these folks need like jerseys or baseball cards. Like I probably should have a trading cards of all of these open eye researchers that I'm trading around, right? That's kind of how it starts to feel when you see some of the huge numbers that these folks are getting paid in these ever-growing bids. Yeah, you know, it's funny.

6:11My kid actually just got his first pack of baseball cards yesterday, which is kind of funny on the time. That is funny. But it does make me wonder, like, what does the rookie card look like for, like, one of these AI researchers, you know? I know I'm definitely the Joker card, that's for sure. but you know it's a great time to be a leading expert in ai clearly our producer adam brought up a really great point like if you're a company that's worth like almost two trillion dollars it's kind of a no-brainer to just spend on this when the potential is there to just completely disrupt the market you know so like if you have that much money like what else are you going to do with it other than spend it all on some ai initiative but really my expectation is that long term this is going to sort of play out similar to the operating system market where you're going to have like a small number of competitors that dominate the space and all have their own sort of unique take on it.

7:01You know, I think about like the difference between Microsoft Windows versus Apple's operating systems and then all of the stuff that's out there that's Linux based. Those are sort of the three areas that the OS market has consolidated around. It wouldn't surprise me to see the AI market do something similar and everyone's scrambling to be one of those three. So it's definitely heating up. Another AI news story. I think this is one of the last ones in our bunch here. An evolution of Airtable to becoming an AI native app. This is something that they announced last week. Airtable, as everyone knows, is a place where you can go store your data.

7:34It's a great way of working with rich, structured data, otherwise organizing things for your company and your team and your projects. But that makes it a perfect environment for tools like Agentic AI that can look at that data and build experiences, tools, forms, and platforms that you can use on top of it. So Airtable announced a app building tool and app generator effectively called Omni. And this is something that allows folks to create interfaces for the data that they store on Airtable. Personally, myself, when I've experimented with these tools, they definitely fall in a different category from like the coding agents we talked about moments ago.

8:09This is kind of like a level higher up on the abstraction in terms of what they are creating for you. And I'm not totally bought in on their usefulness yet. I haven't had a lot of luck generating things from these kinds of apps that actually solved a problem for me that I felt like I could use over and over. I find the most luck actually in applying these things towards automation, like really bare bones, simple automation as opposed to apps. But I'm interested to try it out. I just haven't quite found my footing with these types of tools yet. Yeah. And I mean, we've covered here in the past how like project management, planning, like that kind of stuff really hasn't had a lot of presence of AI yet.

8:44So when you think about the software delivery lifecycle, there's all these different stages. Most of the tools that are out there help you with coding and understanding technical documentation and producing tests and documentation. The space that Airtable operates in, that project management type tool, really has not been penetrated by AI yet. And I think there's a ton of potential there, actually. And I love Airtable, too. We've used it quite a bit. It can be a difficult tool to use at times because it's so complex. but the complexity also makes it really powerful as well. And I actually think that they are pretty well positioned to be like, like, you know, I've been waiting for an AI native project management tool.

9:25I think they are fairly well positioned to be a leader in that, just given the way their platform is structured to help you build data sets that can be used to train AI models and workflows and stuff like that. I think there's a lot of potential there. So I'm definitely going to be checking it out at some point. And for this last story, I wanted to actually flip the script a little bit. This is something that I covered over on the Linear B blog, but it's some new research from Google where they pulled together an experts or a group of AI driven software engineering experts to sort of evaluate how AI is transforming software engineering and really focusing on how there's this widening gap between people who are casual users or who aren't users at all.

10:07and the people who are starting to master the full potential of AI. So they pulled together this group of, I think it was about 21 experts in software engineering, and they defined a set of 12 core goals for AI-driven engineering excellence across the entire software development lifecycle. And that's actually one of the reasons I love this research is because it actually matches our own recent research where we took a look at, we counted 13 steps within the SDLC and to look at where AI was being adopted. You know, I just referenced this on the story about Airtable and we'll have a link to the show in the show notes to this as well.

10:46So you can find it. But they brought together a group of developers who were self-described as being or as AI being transformative to their practices. And they also indicated that they were using AI for use cases that haven't been documented in research or in literature, you know, like novel, interesting use cases that are ahead of the curve. So these experts came together, they defined this group of 12 core goals for AI engineering. They also came up with a six step workflow that these experts were applying to every step within those 12 steps. And really what they're trying to argue is that to get ready for the future, engineers are going to need T-shaped skills.

11:27Like, you know, you've probably heard like a T-shaped person. I think a lot of people sort of strive to do that in many different parts of their professional lives. But really, the essence of it is that you want to combine deep technical expertise with a broader knowledge and context of adjacent domains. So AI will help developers do both of those things. So right now, this research says things like leaders need to help support junior engineers in particular and invest in upskilling because they say that when AI becomes commonplace, today's junior engineers will have to behave more like today's senior engineers.

12:05But then they also talk about how you need to embrace like expanded planning cycles. If AI is producing a lot of the hard work for you, that gives your developers more time to focus on higher order functions. And then also they mentioned that soft skills are going to be really important too. There's going to potentially be a lot more collaboration that happens as a result of using AI. So I think the whole point of this research is that the people and the organizations and the teams who are systematically trying to master AI driven development are going to have a long term competitive advantage, simply because they can make better decisions and move faster.

12:45And I love this research, because it's, you know, with how fast everything is changing, it's hard to keep up sometimes, like we just covered a whole bunch of new features that are coming out. This is something that can sort of help ground you and give you that long-term vision about where this technology is going and what we all need to be doing today to get prepared for it. It's great research and I'm really glad we were able to cover it here and kind of go over what it means for the skill set that folks need. I really like how you called out that tomorrow's junior engineers will be expected to act like today's senior engineers.

13:16That's that decision manager upgrade that we all need to make. The truth is a lot of the folks that are coming in now in the software engineering world, I think that they're really well equipped to use these tools and to come in and have a huge amount of impact really early. The folks that I think that definitely have a lot of work and I find myself in that group is those who have been in their career for a little while and all of the goalposts are about to shift. And so you need to be running, moving every day and learning new things. And this is a great call out of how all the new blocks are falling into place.

13:49So give it a read and figure out where your part plays in that SDLC of the future. Yeah. You know, we have the first generation of like AI native professionals hitting the market right now, like people coming out of college that built AI practices into their work before they entered the professional market. And they're just now, like just now these people are starting to hit the professional world in five years. Those that type of person could be commonplace. And then the people who have been here for a while, who haven't adapted to these new practices, it may suddenly look like they're in a completely new environment that they're not equipped for.

14:25So it's really important to consider those types of transitions. So Andrew, tell us about our guest today. Yeah, speaking of keeping up with new things, we all know about MCP at this point, but are you using it yet? Up next, I'm sitting down with the CTO and co-founder of an AI native company to learn why they're building a first-of-its-kind agency to deliver customized MCP servers. Check this one out. You're going to learn how the future is built. Sounds great. Can't wait to learn.

15:17reviews, head to LinearB.io to get started. Joining us today is Andrew Hamilton, co-founder and CTO of Layer, a company defining the frontier of agent-accessible tooling. Today, we're diving into something that's going to be redefining how your products get used, whether you're ready for it or not. AI agents, they're not just helping developers anymore, they're becoming AI consumers with new protocols like MCP. And we've been discussing this a bit on Dev Interrupted. We had a recent chat with our guest, Asagar Bachu of Speakeasy, who dove into this topic about MCP and how they bridge into APIs.

15:56And that's why I'm really excited to have Andrew here today, a fellow Andrew, to continue informing us about how MCP is going to impact the work we do every day. Because when MCP is on the scene, if you're still building for humans, then you're missing half or even more of your potential consumer base. It's a whole new wave of product adoption. So Andrew, welcome to the show. Yeah, thanks for welcoming me. So we have a lot of cool stuff to cover. I'm going to go ahead and jump into our first one here. You know, we've been talking about MCP, the model context protocol, about how it's enabling LLMs to consume APIs.

16:30So Andrew, you're using MCP to build a whole new layer, no pun intended, for accessing tech on the internet you know why do you think mcp is so transformative of a shift and how products will be used and consumed we've been in the in something that we call the llm extensibility space for a long time now probably year year and a half that's a relatively long time in the ai cycle of thing yes and what we found is that everybody's tried to create this may i call it an app store for llms over and over and over again. The first iteration of it that I saw was ChatGPT with their GPTs. That unfortunately didn't go very well.

17:12There was, after that, GitHub Copilot tried something with Copilot extensions. There were a bunch of other attempts as well. I believe Langchain tried an extendability protocol as well. But MCP is that first attempt at really creating that app store for LLMs, basically enabling you to plug and play your own custom software with an existing client application. And so when you move consumption into this model, you're building MCPs, servers that allow LLMs to make calls into someone's API. Then that kind of reevaluates how you have to package your products, right? Because now they're being consumed in a new way.

17:48Yeah, you know, it's really interesting. It's a user experience shift. And so we've done this with a lot of companies now. We, a bunch of different companies. some of it, the MCP model fits really well with. Okay, they generally have really good iterative workflows. And so the user is aware of what they're asking. And then a change is made, and you're able to see it really quickly. There are some companies where if you just map the API directly, it's not a very good or useful MCP server. I've spent a great deal of time trying to differentiate what makes one product really good for MCP and what makes a product not good for MCP.

18:26And what I've found is that it comes down to the workflows of execution that a user on your platform experiences. A good example of an MCP server that I saw that I thought was pretty neat was centuries. And what it's able to do is it can look at your code, tell you what errors came from the code, hit the API to update or get more information from Sentry, and then update code respectively. And so it's a quick iterative loop. And I think a lot of MCP is finding those types of loops. So it's important to call out an important distinction you made about it's not just about one-to-one matching your API to an MCP server.

19:16It's about thinking about what are the actual flows that people are going to be trying to use your tools in that way. And you call it a good one of using Sentry, you know, it's probably doing combination of like, it's doing analysis of your code, right? But security of its quality, and it's checking it against Sentry's own proprietary information and APIs. And effectively, MCP becomes a way of getting a bunch of really valuable context on demand into the conversation. So maybe that becomes a marker of a good candidate for MCP when it's a really rich context, right? Yeah, the rich context definitely plays a role.

19:53I can give you another great example. The one that I personally use the most is Docker. Very similar flow. One of the first things that happens when I'm writing code is I'll take a bunch of the output of a Docker log, take any errors that it throws me and I'll throw it into chat GPT like immediately or throw it into whatever service I'm using. And that's the first step. And so it just does that automatically, right? It's a tiny little optimization that benefits my workflow. And that's where MCP is kind of pushing the boundary. I have seen some poor uses for MCP solvers. I don't want to mention any or call any specifically.

20:31But there's a lot where they do just directly map their API. Okay, it's like one to one. Yeah, that's what I've kind of seen. So people get confused, right? About like MCP must just be my API and it becomes a bad starting place. what's the point of mcp if it's just going to map your api and i think that's a big problem that a lot of people are bumping into is that oftentimes the best mcp servers use like three or four api endpoints okay but a lot of people map their entire api so for example like twilio maps 1400 api endpoints and that's a lot of endpoints for a model to handle it's a lot of endpoints for a human to handle.

21:14And so, you know, the utility you can get out of a one-to-one mapping just doesn't seem to be where MCP servers really shine. It makes sense to me. It also kind of then is putting a lot of burden on the LLM, not only in its context, understanding this massive library of tools available to it, but then also to understanding how to connect them together, like atomically and to like, I'm going to build, build, build these few API calls to achieve this goal. When what you're calling out is that the MCP server should actually come to the LLM with that chain of API calls or that specific layer of endpoints kind of prepackaged and ready so that it's like a more end-to-end experience instead of it being like a bunch of Lego pieces.

22:01It's in fact just like partially assembled pieces of what they want to build. I actually like the Lego pieces analogy a lot, right? Right. Say you have a few blocks and each of these blocks is a respective API endpoint function or a local function because we don't have to restrict ourselves to just APIs. Yeah. And either you can show the LLM, look, this is a common workflow that our users take, which is an assembled part of your little Lego creation, or you can give it the blocks completely on its own. some things work phenomenal in the use case where you just give it whatever to do sometimes it's really intuitive and really really good really well thought out apis oftentimes that works sufficiently but more complicated bigger apis where you're you have to do a lot of stuff i think we see that workflowing building larger initial chunks for the llm is what allows it to give a better user experience.

23:05That makes sense. And actually, this is starting to bridge into something that you and I have talked about a bit and that I'm seeing emerge as a new topic. And that's this idea of, you know, we have the developer experience, very tried and true part of building any great product is having a delightful experience for your developers and your developer users to be able to interact with the technology you build. But it sounds to me like there's like a similar emerging concept for AI, and you could almost call it like an agent experience. And it's not just like a semantic differentiation. There's actually a lot that goes into thinking about how an LLM would use your API, just like what you called out in terms of what's available to it.

23:48So have you all been exploring agent experience at Layer? And what has that looked like? Yeah, I think a good place to start with that is why do we need agent experience, right? So for example, I mean, this is a straw man, but why is it that you are interviewing me rather than an agent? Why am I even on the interview if an agent can handle it? And I think that that highlights a really good glaring discrepancy between a developer experience, which is designed for a human being, and an agent experience, which is designed for an LLM. I think early on in this, a lot of people had the expectation that agents were going to be just as autonomous as humans.

24:32And so far, we've been pushing the boundaries and it's pretty impressive what we've been able to do. But we haven't gotten to that human level of autonomy needed yet to just essentially use the developer experience. I can't say to an agent, and I'm going to define agent in a little bit here so we are all kind of talking about the same thing. yeah you can't just say to an agent oh onboard me onto this like ai platform and so i think that's where the emergence of this term agent experience has come from so far do you want to define agent for us then so we're all on the same page with what that might look like yeah so when i'm going to refer to agent here i'm just meaning an llm in a loop that has the ability to make decisions as to whether or not it continues working or it terminates.

25:21Pretty simple. It can call tools. It can do that basic kind of stuff. The cursor agent is probably the closest like implementation of that abstraction there. So I think we've seen this a lot in a lot of open source frameworks out there. There's a few that I'll call out. There's Langchain, which has an agent orchestration system, Crew AI. There is MCP Agent, which is one that popped up specifically about, It was by like Last Mile AI. I think they built it. That's a good one. And so a lot of people are taking that initial, you know, do everything agent and constraining it and allowing it to execute specific, very predictable workflows.

26:04And there's kind of a scale from a very weak workflow where you're not really giving the LLM very much direction, but you're giving it a ton of flexibility. so the LLM could do whatever it wants to a very strict workflow where essentially it's not really doing anything more impressive than like clicking through a bunch of buttons. We almost call that RPA, Robotic Process Automation. The company UiPath is responsible for something like that. It's very like set flow. Like it does this one thing and it's just equipped with the tools to do that one thing. Yeah. So it's able to, stuff like that, It's used like UiPath is used very, very frequently at big organizations to automate stuff like filing things or creating a bunch of highly easily automatable stuff.