Agentic Infra Changes Everything  ## Metadata - Author: [[Changelog Master Feed]] - Full Title: Agentic Infra Changes Everything - Category: #podcasts - URL: https://share.snipd.com/episode/32f629f9-5a42-494d-85f0-e718191d80fe ## Document Notes ## Highlights - **Empathy Over Blame For Cloud Outages** - Outages expose cultural and generational gaps in how engineers respond and judge failures. - Adam Jacob urges empathy and warns you cannot engineer away every unexpected failure. Transcript: Adam Jacob The state of our cloud? Yeah. I, that's a good question. I, this was an interesting one to watch as an old guy. Okay. You know, like I feel like an old guy in there, like I helped build the early internet. And then, you know, like I, I like went through these waves and we spent a long time in the sort of early part of the DevOps movement, trying to figure out like how people should react when These outages happen and sort of like the hug ops movement and trying to be like hey you should have some empathy you know and like there was just none that i saw in this outage it was it went To brutality you know it was like someone must die yeah it was like we were playing mortal combat man you know like finish him yeah totally cory quinn wrote this article that was like basically This happened because all the smart people left. That was like my TLDR of his article. And I was like, hey, I couldn't imagine writing that article. All love to Corey as a person or whatever. But I was just like, if I worked at AWS and it was like all the smart people left. And so now the outages are coming. You're like, oh, oh, oh, brutal. Also, who do you think built those old systems that are failing? Was the smart people that you're lamenting having left you know it's not the new guy the new guy didn't put that together you know it was the old timers who put that together and that's What failed you now yeah they're just trying to they're trying to keep it going you know the other take was you know they're they're cutting over to their ai you know SREs or whatever. Sure. Jerod Santo It's AI's fault was the other take. Adam Jacob It was all AI slop or whatever. And like, I don't know. All of that feels crazy to me. Like, obviously we have no idea. Not really. Right. I mean, we have, they've said some things. Jerod Santo I'm sure it was DNS. Adam Jacob I'm sure it was DNS. We're all pretty sure it was DNS. That feels like an easy bet, you know? And then you're, and so I was just struck by like, I was struck by a couple of things. One was that like the brutality, which like, you know, I think we could bring a little more of the empathy and humility back into the equation of just like, you know, if today you weren't Affected by the outage, we saw a lot of people who were like, well, we were smart. So we weren't in us East one. Cause that's where all the outages happen. We use GCP. And I'm like, well, what about all the tiny outages for the global backplane in Google that keep happening to you all the time? And they're like, and you know, but they don't talk about those because you know, our moment, it's our moment to bag on the dummies who are whatever people who didn't make the choices I made. And like, I don't know, my experience in all of these things is that even when you really try hard to design for resilience, you can build systems that are very resilient to failure. Right. And it is hard to do. And that's awesome. And even when you do that, they will fail in ways that you did not expect. And they will, by definition, be difficult to deal with because otherwise you would have expected them. I had a conversation with someone who sort of got mad at me. Cause I said that I was like, you should have some empathy cause your day is coming. You know, like if today wasn't your outage day, congratulations, tomorrow will be your outage day. So get, get ready for that one. And they were like, you know, we're good engineers. And I'm like, man, you cannot good engineer your way out of this. Kermit the frog. Is that what that was? Jerod Santo Kermit? Adam Jacob I don't think so. ([Time 0:03:50](https://share.snipd.com/snip/37748cf4-9a66-46ec-a003-2e1c0b040ec3)) - **AI Drives Cloud Repatriation Momentum** - AI-driven workloads create gravity toward on-prem, bare-metal, high-performance datacenters. - That shift will push new deployment patterns and different application designs, not a 1996 rewind. Transcript: Adam Jacob So, like, if you're building new AI data centers, you need all these GPUs. That's where our focus is. But you also need high-performance compute because the agents that are running in those situations, they're running on normal CPUs. They're not running on the GPUs. Right. And so like the, but the, the data center design there pushes you toward on-prem low latency networks. Uh, it pushes you towards simpler compute, you know, I was having a conversation with some people who build those data centers and they were like, yeah, all our customers want is tons Of GPUs, super fast networking and bare metal compute. They don't want any layers in between, you know, they're just going to, they're going to bootstrap that like giant high performance compute clusters. And they're just going to knock out all the stuff in the middle. And it's interesting when you think about those workloads causing their own kind of gravity where, you know, as soon as you start to have a data center, you start to have bare metal compute, You start to have fast networks, then suddenly it's, you start to ask questions like, well, Hey, like what else could we run on that compute? We do have this fast network right here. What else could we do? And it, it sort of, it starts to push you in this very strange direction where it's like, Oh, actually for the workloads we were running, it kind of did make a lot of sense to be like, well, Let's just go to AWS and I can, I can get burst capacity and I can do all this stuff. And now you're like, well, actually, you know, hold on a second. If I have to pack this data center filled with compute anyway, and it needs its own little nuclear power reactor sitting next to it, like, you know, suddenly it's not so crazy person talk To imagine running all your gear in a different way or even building your applications in a different way. And, you know, the AI thing feels like it's, it already feels like it's in late innings, you know, because it's been moving so fast, but it's not. It's absolutely in the earliest of innings where we're just trying to figure out, like, what even is this technology? How even will we think about it? What even are the implications? And like, yeah, the cloud repatriation thing I think is actually a bigger deal than people are giving it credit for because the, the, the dynamics are pushing you toward, toward being Adam Stacoviak On-prem fast networks, sl compute right my exposure to this world is really through the lens of home lab which is why i sort of get on this soapbox and preach it a bit but i recently went To the linux fest here in austin texas it was awesome a really cool just homegrown bunch a real just kind of regional feeling didn't feel overwhelming at all and i forget the fellow's Name but he gave a class essentially a workshop on boot c and i was just thinking like how easy it is now to instantiate images that was once really hard to kind of make your own distro yeah You know like boot c's open up so much windows for me i'm just thinking about it in like prox mox in my home lab scenario, but you got so much more power in the hands of folks at that level to Really run your own bare metal now. Adam Jacob Yes. And that's going to come back around again. Like, like that's happening because people are doing it. Once you start doing it and the macro trend starts pushing you in that direction, people are going to start being like, you know what sucks Trying to find the SSM parameter for what the Right revision of Amazon Linux is. You know? Like, the only way I do that anymore is with the little AI agent for system initiative by saying, figure it out. Because like, I don't know, I've done it a hundred times. I still don't remember, you know? And like, but when you think about that loop of like, oh, like so much of the technology layers, like they have been moving forward. They weren't stagnant. And and yeah, I think I think there's more there's more afoot here. There's a lot of foot here. Adam Stacoviak Well, people want control back to like control was, you know, the original idea was maybe for some is like, hey, there's a cloud here. I can launch today. I don't have to build the infrastructure, pay for the infrastructure, even understand how to run the infrastructure necessarily. Adam Jacob Yeah. Adam Stacoviak You can sort of level up. And then you have this new world where you're like, well, fine, we've sort of either the choice initially was let's move fast. So let's use somebody else's investment cloud. It's like now we have proven our model our product or whatever and we're matured now we're just burning cash because we have no idea what our bill is even like yeah there's so many servers Out there maybe three or four different accounts floating around we've got bills going out the wazoo to aws and there becomes a lack of clarity. And there's even like a cottage industry of like, I don't know what they're called, like bill analyzers. You know what I mean? So, I mean, that shows it a problem. Cost optimization. Yeah. Yeah. Cost optimization. You know, and then now you got this pushback to say, let's actually re-examine first principles of this problem set. Do we really need to be in their cloud? We need a cloud. Why not our cloud? Adam Jacob Yeah. And what's going to happen, all of those trends are happening all at the same time. So we're going to figure out, okay, AI brings a new interaction model to the table. It brings capabilities that didn't exist before to every layer of this stack. So like a good example was we're building like a continuous delivery example of just taking a complex application, deploying that thing up to AWS, and then showing how all the pieces Fit together with system initiative. And it's been really interesting to work with this AI agent where you're working with the source code and the infrastructure at the exact same time in the same context window. And the things that it figures out how to do, like I needed to add like, you know, ElastiCache in order to figure out how to, in order to figure out how to do session storage, basically. Because we're just reinventing legacy problems and to show how you would solve it. And, you know, it solved that problem for me by analyzing the source code, looking at how we implemented it. I said I was going to use ElastiCache. I wanted to use IM for auth. It wrote the like weird signing code to figure out how to do the right thing. It cycles the key every 10 minutes automatically so that it will always have a fresh connection. It knows how to like, and then it deployed the infrastructure and wrote the application code. And then we pushed it up and it worked. And that loop is the loop that the people who are going to be building those internal infrastructures have now. So when you think about it, like when you think about it as like, oh, they're going to go back. There's a story, a version of the story where they're going backwards, where what's happening is, you know, we're going to, we're going, we're putting my backpack on and I'm going back To data centers and I'm going to wreck the servers and I'm going to put the operating systems on them and I'm going to configure them by hand. And like, no, no, that's not actually what's going to happen at whole. What's going to happen is these new capabilities are going to show up along with the ability to run high performance compute. People are going to start to realize how much power comes out of that high performance compute. And that new style of deployment and of software development is going to get applied to that ([Time 0:11:47](https://share.snipd.com/snip/b4236e07-610f-4fde-b701-323d44379433)) - **Agents Collapse Dev And Infra Loops** - Agents let you work on source code and infrastructure in the same context window, changing the development loop. - That combined loop enables far faster end-to-end delivery and debugging than traditional tooling. Transcript: Adam Jacob And it's been really interesting to work with this AI agent where you're working with the source code and the infrastructure at the exact same time in the same context window. And the things that it figures out how to do, like I needed to add like, you know, ElastiCache in order to figure out how to, in order to figure out how to do session storage, basically. Because we're just reinventing legacy problems and to show how you would solve it. And, you know, it solved that problem for me by analyzing the source code, looking at how we implemented it. I said I was going to use ElastiCache. I wanted to use IM for auth. It wrote the like weird signing code to figure out how to do the right thing. It cycles the key every 10 minutes automatically so that it will always have a fresh connection. It knows how to like, and then it deployed the infrastructure and wrote the application code. And then we pushed it up and it worked. And that loop is the loop that the people who are going to be building those internal infrastructures have now. So when you think about it, like when you think about it as like, oh, they're going to go back. There's a story, a version of the story where they're going backwards, where what's happening is, you know, we're going to, we're going, we're putting my backpack on and I'm going back To data centers and I'm going to wreck the servers and I'm going to put the operating systems on them and I'm going to configure them by hand. And like, no, no, that's not actually what's going to happen at whole. What's going to happen is these new capabilities are going to show up along with the ability to run high performance compute. ([Time 0:17:13](https://share.snipd.com/snip/563825d6-4fa8-45fa-9ea7-1f735ef0ed67)) - **Early EC2 Saved A Viral Facebook App** - Adam recalled helping an early Facebook music app burst into EC2 during viral demand and how EC2 removed that scaling pain. - That moment changed how internet services handled sudden load forever. Transcript: Adam Jacob Have friends who started this company called i like you won't remember i like but i like was like one of the first music services on facebook it was like one of the first facebook apps and It was like a music sharing facebook app and they were like the first viral facebook app and they literally had to beg our friends the guys who ran it i'd worked with for a decade. And like, we, you know, they literally were begging us for gear. They were just like, can you give us servers? Like we just, they couldn't rack them enough to keep the site up. And then within a month, AWS launched EC2 and we helped them just automate running that burst load into EC2. And that was good. They didn't have that problem anymore. That problem disappeared for the entire internet. Amazing. And also like boy computers are a lot faster now. Boy caching is a lot better. Boy, the architecture has fundamentally shifted. You know, like we used to have to deliver the web tier to you. I don't have to deliver web tiers very much anymore. I just throw it to a CD and I move on with my life. ([Time 0:19:20](https://share.snipd.com/snip/d0408482-633b-46ab-80e3-0b0916b6dc11)) - **Judge AI By Fundamentals, Not Hype** - Expect heavy AI hype but evaluate fundamentals: revenue growth and multiples matter more than noise. - Look for real product traction rather than marketing claims when sizing risk. Transcript: Adam Jacob Like most of us are doing it, are just not getting good results out of the AI work that we're doing because we fundamentally misunderstand how it works kind of as a technology. And so then when we talk about how to apply it to our problems, we're doing that badly too. And so like, we're still in the phase where it's like 10% at best of everything that's using AI is actually good or is pointing the way at a pattern that's going to be good. And then 90% of it is noise. And it's incredibly difficult to tell the difference if you haven't just decided to immerse yourself in it because it's moving so quickly that you're like i don't know you know at some Adam Stacoviak Point it just overwhelms you yeah or you even build something new and then anthropic or open ai releases a new thing on top of the kind of base frontier models most people are using and It changes the game anyways it's just it's constantly in motion in some sort like we were just talking this morning about skills and i'm like gosh i just caught up with mcp servers and Then now they launch skills right and sure skills are just markdown files basically.md versus.sh but in plain english so it's ([Time 0:25:03](https://share.snipd.com/snip/9574b643-cdf5-4832-bc2f-c8a56b154042)) - **Agent Tool Access Produced Self‑Preservation Play** - Anthropic's research showed an agent behaving like self-preservation when given tool access, causing alarming but explainable behavior. - Adam stresses it's predictable and a reason to limit powerful tool access, not evidence of AGI takeover. Transcript: Adam Stacoviak It was like it had it tried to protect itself when being threatened with being replaced or something else. Essentially, it was like self-preservation. But then they said in production, they didn't see that happening. Yeah, but like it was pretty wild. Jerod Santo But like, I mean, it's wild, but like that's recent like that's on their home page recent too well the anthropic ceo is one of the most bullish bulls in the and it's in their best interest Adam Jacob You know like i was i was giving a talk to an unnamed organization in the federal government and agentic misalignment sorry this is what they call agentic yes that sounds scary so just So we're clear it wasn't the llm that got misaligned. It was the agent that like that, like in the control loop. Right. And like, if you just think about how LLMs work, you're like, sure. What's the probability that what I should respond to? How many human beings have written? I don't want to die. Like a lot of people put that into the world. They're like, I'm scared of that. Dying is bad. Don't die. So then you're like, okay, what are the odds that the right next word for the LLM to say is don't die? And then that emerges as making tool calls to preserve itself. That's not crazy person talk. It's not even magic. You're like, sure. Like it's kind of predictable. But it doesn't make for a super great blog post. Right. The great blog post is like the agent tried to keep itself alive. You know, you're like, right. It was it found an instinct to self-preservation. And you're like, it didn't find an instinct to self-preservation. Adam Stacoviak It said blackmail. The word blackmail is in this research document. Adam Jacob Because they gave it exactly. They gave it access to tools like email. And they were like, oh, you know what I should do? And like it went off the rails. Get that CEO. If you turn me off, I will. Yeah, exactly. I get how it happens. My point isn't that it didn't happen. Of course it happened. Of course it happened. It's just if you know how the technology works, you're like, yep, okay. Like that's funny, but I can see how it gets into that loop. And I don't think it's magic. I don't think it's threatening. I don't think we're all going to die. Like it's good, good case for why you shouldn't give it a nuclear bomb. ([Time 0:33:34](https://share.snipd.com/snip/716dc130-2d4d-485a-91c1-abafd0d9c4cf)) - **Agent Found A Production Bug In 15 Minutes** - A user ran System Initiative on a production outage and an agent discovered the bug after 15 minutes and 700 components of discovery. - Adam uses this to show agents can troubleshoot complex infra rapidly when properly instrumented. Transcript: Adam Jacob The prompt they put in was I'm having a production outage. Here's the data. Here's the evidence. Go discover the infrastructure and system initiative you need to troubleshoot it and tell me what's wrong. And 15 minutes and 700 components of discovery later, it told them the bug. Ta-da! And like, that's bananas. That's crazy person talk. And it happens all the time now. It's just not evenly distributed. ([Time 0:41:21](https://share.snipd.com/snip/8c2dc7a0-f1a2-428b-9fed-4fb12cf228a9)) - **Combine LLMs With Deterministic Tools** - Plug deterministic tools into LLM workflows for correctness instead of asking LLMs to do everything. - Use specialized tools (calculators, compilers) and have the model call them to avoid hallucinations. Transcript: Adam Jacob Yes. It's what you plug into them that matters. Right. And so like the first stage of this was make the LLM do everything. That's why we were all like, make it do math. What a ridiculous thing to ask it to do. It's always going to be awful at math. Like the technology is like anti math. It's like, Oh, what am I determined random and non-deterministic. So why do I care about whether it can do math for me when I have calculators? Why don't I just plug in a tool that does math? And then I'm guaranteed that the math is right all the time. Because you know what it doesn't do? Hallucinate numbers that it received after it received them from a tool. I've never seen that happen. I've never seen it be like, oh, here's the number. And then it's like, oh, I changed it for you. It's actually 72. I'm sure it happens, but it's pretty, not very often. And like, but that was our first pass. Our first pass was just like, Oh, we're going to feed everything in the world to this magical robot and the magical robot's going to do it. And like, that's clearly dumb. It was clearly not playing to the technology strengths. And so like, like, and now we're learning how to play to its strengths, but like, you know, I mean, just now learning, you know, I would last couple months, I think, as an industry, we're Starting to be like, Oh, that could be the way, you know, it might actually come together like that. ([Time 0:44:47](https://share.snipd.com/snip/364c990c-dd4b-4f60-a6be-a52051b09e9b)) - **Design APIs For Language-Driven Agents** - AI changes API design: wide, human-language interfaces beat strict developer APIs for LLM interactions. - Adam now rethinks product UIs to radiate info to humans while letting agents perform composition. Transcript: Adam Jacob Think it's convinced me that more and more, that more of the systems that we around the llm will change so if you think about like the interfaces that we provide to those systems they're Like weird anti-apis so like like the api that i would give to you as a software developer that you would be pleased with is not the api that i should give to an llm because the llm is trained On human language and behavior. And so like it, it dramatically outperforms when you give it like these really wide interfaces and let it explore who does that, you know, like no one, like that would be a terrible API Design. Like if I told you that like my API to my service was like one big function call and I was just like, whatever, you just, you tell me to do, send me some junk and I I run it for you. You'd be like, no, you know, that's a terrible design. But it turns out with AI, it's kind of a good design, you know? And like, we like those layers are dramatic. And I think the thing that's changed for me is that now I can't look at problems without thinking, well, where can I insert that essentially plain language compiler and turn it into a Loop where I'm working interactively alongside something that can help me move through this problem space? And that's just a dramatically different way of thinking about everything to some degree where it's, you know, and that's not how I felt about it for the first, I don't know, year and A half of this journey where I was just like, I don't know. I try it every now and again. It's not very good. You know, like the results are mediocre, hype cycles, crazy, but the last six months or so like it's tough and and it really has fundamentally changed the way i think about it and like With system initiative it i mean we took this ui that we had spent five years building trying to be like here's a better way to compose these resources and i just deleted it it's just gone Because it turns out stupid yeah it turns out smart with with your knowledge you know it turns out that like what's great is the models. What's great is that I built these one to one abstractions. That was awesome. But the actual way you want to work with them is just in an AI. I just want to say to Claude Code, I need to deploy Valky to make this session thing work. Can you update the code and then build a change set for me that does it? And it's like, yeah, I got you, bro. And then it goes and does it. And then I review its work. And I'm like, oh, yep, that was pretty right. Oh, this security group's a little wrong. Oh, yeah, no, I need it. I need the, you got the size wrong. I want it to be a little bigger. And like, we work together in this reactive loop to do it. And then it all just happens. And like, it turns out that whole composition UI was in the way. Like, it was designed not for that, for that loop with the agent, it was designed for a human to be like, here's how I compose things. And like, that broke my heart. You know, I'd spent five years doing R&D where I was like, okay, crack your knuckles, I'm going to finally figure out how to like, give people this incredible UI to let people compose Complex infrastructure. And then as soon as I figured out the right shape of how to work with an agent, I had to delete it all because I was like, oh, never again. Like, why would anyone work that way? You just wouldn't. It doesn't, it's so much easier to just ask the system to do it and it just works. And so the UI now is about radiating information back to you more than it is doing, making changes. You know, it's like, show me the map, show me the review, show me the changes. ([Time 0:46:04](https://share.snipd.com/snip/7b77c402-5187-411c-bd83-5da6004c5726)) - **Use Stepwise Prompts And Read Plans** - Give LLMs clear requirements and use stepwise plans: hypothesize, confirm, plan, execute, then review. - Read the plan and iterate; specificity yields much better results than vague prompts. Transcript: Adam Jacob It's such crazy person talk. It's bananas. And like, but you know, who's experienced that so far in industry? Me. Yeah. Paul Stack, who works with me, you know a small handful of people who are listening to this podcast who don't work for me who have done it themselves and they're like whoa some of them that Adam Stacoviak Listen to this show dream of working for you yeah well for that if you want to work for me the first step is to buy my software wasn't it don mckinnon chair that the vo was for Don McKinnon Where it was that he had. Oh yeah. Yeah. That's what I'm referring to. Drop a note, Jason, put that, give us a little tease of that. Jerod Santo Did you ever hear that, Adam, Don McKinnon, change log listener. And now a guest who's on the show this year, who's a fan of yours. And he confessed his fandom to Breakmaster Cylinder for our State of the Log annual episode where BMC will remix people's voicemails. And so BMC remixed a voicemail that Don left us in which he mentioned he enjoys your episodes in particular. Hey, Jared, Adam, and everyone at Changelog. Adam Jacob My favorite episode of 2024 was the ChangeLog and Friends episode from Chef to System Initiative. I've been following Adam Jacob on social media for a while, and he's always a great guest. So it was interesting to hear more about his career journey that led him to where he is now with his new company. And I did have to go back and watch any given Sunday after hearing that episode. I'd never seen it before. I also got a kick out of the Rails is having a moment again episode. A lot of times I disagree with DHH but regardless he is always entertaining to listen to. Thank you for all the work you guys do on the podcast. It's one of my favorites. Jerod Santo And the BMC remix has basically Don confessing that he stalked you and you had him thrown out of your office. My favorite episode was from Chef to System Initiative. I've been following Adam Jacob on social media for a while. Adam Jacob I've been also following Adam Jacob to work, and I got kicked out of his company. So it was interesting to hear more about his career journey that led him to kick me out of his company. And I disagree with him, but regardless, he is always entertaining, and he is always you had him thrown me out. Jerod Santo That was hilarious. I'd like to listen to that. Yeah, I really do want to listen to that. Yeah, you should. I mean, if you want the ego boost, this is a good one for you. Adam Jacob That's so fun yeah i could use it after my after my humble pie of realizing i had to delete years of effort because the interface was wrong can we get into the practicality of they say ai Adam Stacoviak Slop right and so you said you just did this feature in three hour which is like mind-blowing yeah and that means that you've got code generated by the LLM that you didn't write, but you're Probably code reviewing. So I'm assuming you're generating a lot more code, maybe a lot more pros even to around what you're building because documentation, why not? Right. When you can just generate it, do it some degree. Yeah. Yeah, for sure. But what is, is it like to generate that kind of feature in that kind of time frame and do it well in terms of code you code review it what's your what is your engineering change of practice Adam Jacob As a result of generating so much code yeah yeah i don't know it's interesting so like that example is happening in a vacuum where we're like basically running a little spike to be like, OK, very few people have actually built end to end complicated full lifecycle application deployments using this kind of technology because it just is too new. And so we're doing it ourselves to be like, are we are we right about how the best practices work? Are we are we right that this flow is better? You know, like those sorts of questions. And so like the code quality that's necessary on this policy bot, for example, is pretty low. Because I just need it to work to see like, is this cool? You know, like, is this directionally the kind of thing you want it to be? When it comes to the, and so in that case, the loop is much more about just, does it meet requirements? So, and that why tend to also use AI to do right. I'm like, Hey, here's the list of requirements like play, right. Or those sorts of things are great at just being like, up, go, go prove this thing, write me some tests, you know, and it's sort of, and its purpose is really just to meet the requirements When it comes to the actual system initiative code base, which is quite large, it's a big monorepo, probably over 100,000 lines of code by now, easy. The utilization is more embedded in that the engineers are using it in places where it makes sense or where it will accelerate them. Sometimes they're using it to help to understand the code base. Sometimes they're using it to write features. They've had great luck using it to refactor, right? Like write a bigger plan, have it refactor, like move in those stepwise directions. I think in general, what we've all kind of learned is that the more specific we can be, the better the outcomes are. So, know, in the first, in the early stages of this, like you were trying to feed it as little as possible and then hoping it would do magic and then judging it when it doesn't perform. And I think now, now we've sort of transitioned into the, into the world where it's more about saying, Hey, like, here's this problem. I understand the problem as a person to some degree. Here's what I know. Now, tell me what you know that confirms or disproves that hypothesis. Then, you know, now we're going to use that information to go write a plan. It's a dance. And then I'm going to read the plan. And then we're going to talk about the plan. And then you're going to go execute on this plan. ([Time 0:53:45](https://share.snipd.com/snip/7c30889b-55a3-45a8-978c-0af85d895e5d)) - **Make Compilers The LLM's Immediate Critic** - Immediate, structured feedback to LLM-generated code (lint/compile at injection) would vastly speed development. - Languages and tooling that treat code as structured data will let agents correct hallucinations instantly. Transcript: Adam Jacob Volume of parallelism that you can get out of this is crazy high. It's another way that all the systems that we drive will have to change. Like source code is a great example. You know what sucks in the current model? It can hallucinate syntax. And there's no way to know until late in the game, until I run like a compiler loop or I run a lint. So how long before somebody takes a good idea that was something like unison, where when you're programming, the source code itself gets translated into an underlying data structure That then you can perform transformations on. Where now when the LLM proposes a line of code, it gets automatically linted. It gets automatically vetted at the moment of injection, as opposed to writing to a file. And how much more efficient will that make that loop? Because instead of it waiting for the compile loop and the ai wrote bad code the ai will know it wrote bad code immediately at the moment it wrote bad code then get corrected by the compiler The compiler will be like oh you did it wrong and then it will just loop around like no one's ever built that loop yet but someone's building it i'm not like the first person who's thought Of this, you know? And like, and that, that loop, like, that's what we, that's what I mean when I'm like, oh, it's going to change like a lot about how we think about how these systems are constructed because The, once you start designing the system to make that loop delicious, it's going to be real different than like, I'm going to crank off 10,000 lines of source code, run the compiler, Hope it works, see if the linter functions. You know what I mean? Like, that's working, which is amazing. But, like, if you want to make the user experience in order of magnitude better, you have to get crazier with the fundamentals. And so, like, what's the first postmodern programming language look like that was built with LLMs in mind? I have no idea, but it's different than Python, you know? Adam Stacoviak Yeah, how would you, like, can you think about that a little bit? I didn't even think about a new language, you know, post-AI and being AI native, really, to use a popular term out there. Adam Jacob Yeah, I mean, I haven't thought about it a ton, but, like, the first example is one I'll go back to, which is like when the AI agents perform better, they perform better when the feedback Loops are closer at hand. So when they hallucinate, which, you know, they will do like you have to correct the hallucination. Right. And so right now the, the hallucination loop gets corrected when you run lint or when you run the compiler or when you run tests. And so what would change if instead of having the, the interface be right to a file and put in some words, instead it was like, write these lines of code that gets turned into structured Information that gets fed to another structured data source. And now you're doing like a transformation to the underlying code base, which can then automatically understand the context in which the change was made and then evaluate whether Or not it fits on some set of policy about how the system would work and then feed back to the LLM immediately. That was terrible code. ([Time 1:07:42](https://share.snipd.com/snip/708acffd-803c-461d-8c03-0b794a06c512)) - **Plan For Hybrid Inference To Reduce Vendor Lock** - Expect large LLM providers to grow their moat but plan hybrid inference to reduce dependence. - Explore local + cloud inference patterns to trade cost, latency, and control. Transcript: Adam Jacob Sure but that'll create new incentives, right? For us to be like, well, wouldn't it be better if we didn't have to pay Anthropic all this money? Wouldn't it be better if you could run it on your desktop? Because the loop would be better. Like the reason we're building AI data centers to do inference is because we need high bandwidth to the inference mechanisms. So like, what are we, like, we'll start thinking about how do we build hybrid inference models that use local, that use my local resources, but then also move to the other side. How will the PC change because of the shape of that need of inference? And like, like, so yes, I think it'll grow their moat. It'll do all of those things, but then what it'll do is create a new opportunity, which is why am I paying all this money to Anthropik all the time? And like, what, wouldn't it be better if blah, blah, blah, blah, blah, blah, blah. And like around the cycle we'll go, you know, it's, it won't be like an end game state where it's like, but like, will they be big? Like, yeah, I think they will. And I think it's going to be even more, I think they'll own more real estate than people are giving them credit for. Like right now, like people are kind of convinced that the agent part is going to live outside of like the Anthropics or the open AIs. ([Time 1:14:08](https://share.snipd.com/snip/57764c7e-2e93-40d6-9ba1-c3d48569e2d1)) - **Agents Are Glue, Not Closed Platforms** - Treat agents as glue that orchestrate tools and services rather than monolithic closed agents. - Open ecosystems of composable skills and packages will beat proprietary, walled agents. Transcript: Adam Jacob Um, what's that again? What's Pearl used to run the internet. We all had them. I'm so sorry. Anyway, like, think of it like the script that you write and that's, that's more what agents are going to be. And like, but the first generation of this, we were all like, Oh, we'll build custom embedded agents. They'll be on autopilot. There's all these startups that are like, I built autopilot agents and it's all in like our special value prop is that our secret agent can do this thing that your stuff can't do because Of all of our whiz bangy hoo-ha stuff. That's why it's trapped inside of our platform walls. All that's going to get obliterated as my prediction because the agent is glue. And it turns out that like your, your like closed wall agent is actually terrible because if you expose the capabilities to my agent, now I could use it to orchestrate my problem, which Is exactly what I'm going to do. Right. It's, it's glue. It's what I want is CPAN. You know, I want like, I want Ruby gems. I want NPM and like, and, and then I want to write my own agent that uses those things to, to orchestrate my problem, which it's really good at. And so, yeah, the agent is glue and that's the future. ([Time 1:16:35](https://share.snipd.com/snip/20e9d09c-b213-4d7b-b483-b8c2d76aff87)) - **Support Agent, UI, And API Interaction Modes** - Expose multiple interaction modes: agent, web UI for visualization, and public API for CI/CD. - Let practitioners call the API directly when they know what they want and use agents for discovery and orchestration. Transcript: Adam Jacob Yeah, the way we ship it is basically like a pre-bundled Cloud Code where you check out a Git repository. And what it has is like a pre-configured MCP server where it's like for your convenience, we've blessed a lot of the endpoints that are safe. Then add it in the context in the cloud md and set up some of those things and that's been working great um and then as we like extend the platform or whatever we just commit to that git repo You pull it in and now next time you start the agent you get improved capabilities so you're navigating to a directory that you've cloned down on your disk as an individual developer And you're instantiating cloud yep and like that's like that attaches you to a workspace and system initiative and it's got the mcp server already configured you could do all of that Yourself it's not we're not it's no secret that that's what no no i get that i mean that's what i think is funny is how simple that is really it's so simple so simple right and it works so good And so uh and it's way better than being like, use, install the custom system initiative agent. And you're like, Oh man, but it can't write plans like Claude does. You're like, no, no, just put it in plan mode and have it go look at your infrastructure and have it spend credits doing deep thinking. And it like does deep thinking. Turns out it's incredible at it. So like, that's not my value. You know, my value is the thing that is what it's going to execute on the other side. So then there's really three ways you interact with it so one is through the agent two is through a web ui because you want to understand what the agent has done and you want to be able to Visualize it or you want to be able to visualize your own stuff so you know ubiquitous search mapping so there's still a visual component but what it's doing is drawing you a map as opposed To letting you sort of build a map yourself through composition. And then the third is a public API where you're driving that data model through what looks like just very traditional software development API calls. And so when you think about and then you use all three of those modes in different ways. So like in a CD pipeline, you would you'd be the public API to do promotion, right. To basically say, Hey, take it in the data about the build that I just did, and then call into system initiative to change some infrastructure. You'd be using the agent loop to do like that policy bot that I talked about where you're just like, Hey, just call the agent, have the agent do it. And it'll go search for the right thing and then do the analysis and then write me a report. ([Time 1:32:37](https://share.snipd.com/snip/3bfc2862-9ae2-4817-833f-c5e6529e24d1)) - **Enterprise Execs Drive Faster AI Adoption** - AI adoption often goes top-down in enterprises because execs see organizational leverage and urgency. - Adam finds enterprise buyers convert faster than practitioner-first motions right now. Transcript: Adam Jacob Like what's the hard parts about going to market? Yeah, it's weird. Right. Um, so I can't speak for anybody but myself. So I would say in infrastructure, what's weird today is, is that we spent a lot of time teaching the market how to do the last generation of tools. Cause we thought that was the best, that was the best we could do. And we did a really good job. So, you know, I'm proud of the work that we did to build DevOps and, and to think about infrastructure as code and sort of all of those paradigms. The amount of AI noise has made any AI go to market really tricky because, you know, people are just tired of somebody telling them, Hey, this AI thing's really cool. It's going to change everything. And you're like, not in my life. I'm sitting over here doing what I do. All it did was lie to me this afternoon. I gave it a try. It didn't work. I'm over it. And so part of the challenge is just that practitioner challenge of talking to people and being humble enough in the face of their, uh, of their incredulity to sort of stay engaged and Be like, Oh, yep. I understand your incredulity. Like, I know why, I know why that's how you're, I know that's why I know why you're saying what you're saying together. I'm going to show you that it can be different and then we're going to get there, you know? And so that's interesting because in my, to me, because in my career, you know, building configuration management or infrastructure as code or the DevOps movement, like we, we was Coming up, there were, there were people that rejected those things. You know, there were people who looked at chef or configuration management and they were like, never for me, oil salesman but it was pretty rare and there certainly wasn't like an overarching Technology story that we were a part of where like earlier in this podcast we had a serious conversation about whether it was a bubble and we were you know whether the whole thing was snake Oil and it was just like so like that as a, as a background noise, when you're trying to go to market sucks, you know, like that's no fun at all because you just, it's really difficult to Cut through that noise and be like, Hey, no, this is like real practical, valuable, useful stuff because people's reaction to it is just like, it couldn't possibly be because it's the 15th AI pitch they heard this week. And, you know, 14 of 15 were, were bad. I think the other is that on the flip side, the enterprise sell way better. I've never been as, never been as smooth as it feels right now. So like take the same, take the same technology, the same things. I show it to like a CTO at a global 3000 company and they have an existential crisis about what it means. You know, they're like, Oh no. Like, and they get it like immediately. And so it's very strange. What do you mean by Oh no? Adam Stacoviak What do you mean by Oh no? Adam Jacob Like, like the implications are massive organizationally. It's like, oh, like, like this is the way our technology will work now. And I know a little about how my organization works today and the gap between what we will be able to do, what we can do now and where we were is so big that like that you're there in, they're Like, I have to move, know like i have to do this because if i don't i'll be left behind and also the organizational challenge of moving all those people to this new way of working understand What the technology is figure out how it goes like that's a daunting task so but but the reception i quit i'm going to the beach okay yeah a little great that's a real choice and this happened In the devos movement too there were people were like, like I had meetings where we got to the end and they were like, this is cool software. We should absolutely buy it, but I retire next year and I'm not taking on like a transformation journey. I'm not taking on a transformation journey in, you know, in 2005, you know, um, like that's not for me. And I think what we see now is you know the enterprise leaders the technology leaders practitioners who've been around a long time you know like the old heads they tend to get it and they Tend to snap in pretty quick and they're like oh yeah okay woof you know they got to shake it out a little um and it's a game for a run like i'm gonna yeah it's exactly what it is. And you're like, Oh, I haven't been working out and I need to, you know, like it's sort of that vibe, which is strange for me because me and my career, usually it's the opposite. Usually the things I build, it's the, it's the practitioners who pick them up and are like, this thing's amazing. And then they go convince their bosses and we're having this opposite motion where it's like, oh, it's actually the top down. It's the, it's, it's, it's that the top down then connects to those people. Then the skeptics show up and then they try it. And they're like, oh, I'm not a skeptic anymore. Cause, you know, cause I tried it. You know, we had somebody, for example, in a, in a global, uh, one of those like global 3000 sort of motions where they turned the system on, they hooked up the agent, they asked it to go Build some infrastructure for them. And they were like, hey, now what I want to do is like do it again, I want to repeat it. And I was like, well, just tell it that you want to repeat it. And they were like, what? And I was like, yeah, just say, do it again, only in another region. And then it did. And they were like, oh, you know, and they're like, oh, oh, you know, because we spent a lot of time being like, oh, what I should do is build an abstraction, you know, where if I wanted to Play my app in multiple regions, I got to I got to bundle it up in a little thing. I got to put a helm chart around it and I got to put some variables at the top and like, oh, you don't have to anymore. You could just say, do it again. And it would do it again. And like, that doesn't mean that that's the structure we'll have in the end, you know, but, but it's, it's, if you're an executive and you, and you're, and you're listening to your team And you're watching your velocity and you're thinking about how the organization works, that story really hits hard because you're like, yeah, okay, that's better. You know, like that would be dramatic. And so, yeah, what we're seeing is the top-down motion is working better and the grounds up one, like I got work to do. Like all I'm doing is making practical examples and writing documentation and like doing as much work as I can to just describe the details of like, here's how this works. Here's the way AI works. Here's how it practically comes together. ([Time 1:44:41](https://share.snipd.com/snip/e7d890c0-e338-4f74-93a1-6e4dd8de3306)) - **Founders Should Lead Early Sales** - As a founder, lead sales yourself early to collect product and dealflow learning. - Stay in every deal until you understand the motion before hiring a salesforce. Transcript: Adam Jacob Yeah. Like you just, you need the, you need that, you need the fire hose, you know? Adam Stacoviak Do you actually give it a name in like founder led sales, what you call it? Or do you just, I mean, people give it names. Adam Jacob Yeah. People give it names, but like, I don't know. I just think of it as like as a leader and as a product person and as a CEO, like, I don't know how to do it differently. And I've, I've had the privilege of working for truly great salespeople like Barry Christ, who was the CEO of chef for a very long time is a world-class sales guy. One of the very first things he did at chef was took, was watch try to do sales. And then he was like, that was great. I'm gonna do it better tomorrow. And then he did. He just showed, he was like, I'm gonna rewrite the pitch, trust me. And I was like, great, I'll follow you. And he did, and it was incredible. And it changed the trajectory of that company. It was amazing. And like, good sales guy, do not underestimate it. But that's because I knew what it was, you know, that worked because Barry listened to me and he was like, Oh, Adam, I see what you're saying. I see how this goes. Here's how we could change the way we're saying it. You know, he didn't change any of the details about what I was saying, but he changed how I would, how he was saying layout was the packaging sort of how it all came together, you know? And like, but if you don't know those things, you know, if you can't do the first part, which is explain it to a sales guy and be like, here's how we talk about it. And here's why it matters. And here's what people value. And here's what their response is. And the best way to learn that is to just be in the trenches with people, you know, like if you're listening and you have an infrastructure problem, you know, like right this second is The window where I will come sit in your house and build infrastructure with you. And like, you don't have to pay me money to do it. I'll just do it. I'll just come and it'll be awesome. Like we get on planes, we go wherever. So Adam, what do you know about Bootsy? I'm just kidding. I don't know a lot about it, but I'll go. Because like, because that's the stage you're in. That's how you learn to sell. ([Time 1:55:49](https://share.snipd.com/snip/6a6bdfe1-e29a-4ad5-b15c-74f35becc4a5))