The Agent Meta Shift Most Builders Haven't Caught Yet

Ben Thompson from Stratégie had this article about the differences between 4 deployed engineers from different companies. Jonathan and I just synthesized this in ChatGPT based on open job recs to see from Google, from Palantir, Anthropic, OpenAI, like, how are they thinking about this? What are the heuristics of different FDEs for deployed engineers that they're hiring for, indexing towards? So different heuristics, custom build implementation, AI model depth, enterprise workflow and ops depth, uh, sales and solution orientation. We had discussed that Google is really going for this, hey, just like wire up all their data, throw it into Gemini and we'll figure it out. We'll index it. Palantir takes a much more customer-centric approach. You could see by this green bar. So really understanding the enterprise workflow. That how they do operations. Anthropic has a little bit less, OpenAI a little bit less, um, and they're all taking their own philosophies and actuating on them in this current environment to try to win in the enterprise game. So, uh, another interesting take is, um, from Palantir's JDs, it actually talks about owning customer outcomes. The other ones don't talk about that. They're much more delivery-oriented. So super interesting. And something else that Jonathan had mentioned before this was how the colloquial wisdom now for AI builders such as us and anybody listening is earn the right to solve your customers' problems, get obsessed with them, really build that skill of empathy, over-index on it, and then you earn your— the right to solve more customer problems. So I'm gonna be over-indexing on this green bar, uh, is how I'm thinking about it. Cool. In other AI news, uh, GBrain— Gary Tan, who is the president and CEO of Y Combinator, built GBrain and released this earlier this year. GBrain is just a GitHub repository that bakes in a level of cross-session memory into OpenClaw or Hermes agent deployments. So some stats: 17,888 pages, 4,383 people, 723 companies, 21 cron jobs built in 12 days. That's, uh, that's GBrain. It is a self-wiring knowledge graph, and every page write extracts entity references and creates typed links. So for Gary, it was stuff like attended this meeting, works at there, invested in X, founded this company, advises who. And it does all that with zero LM calls. Um, there's a level of hybrid search that backs things for backlink boosting, uh, of ratings, or I should say rankings for search rankings. And he had an interesting benchmark. Um, yeah, here we go. Benchmark side by side against the category. GBrain lands P@5, 49.1%, R@5, 97.9% on a 240-page opus-generated rich prose corpus, beating its own graph-disabled variant by 31.4 points. Cool. Interesting. Um, thoughts on GBrain? You mentioned you had a friend that was using this and was hyping it up. I haven't used it myself, so, uh, what did you hear and what do you think of it?

Yeah, well, I actually don't know anybody that's, that's using it. He was excited to like play with it over last weekend. I'll check in with them, see, see how it went. Um, and, uh, yeah, I think like this is an extension of OpenClaw and Hermes usage. Um, I love that Gary Tan is using GStack to then like implement GBrain, and that's all part of the, this sort of like offering here. So that's pretty cool that you have somebody, um, especially somebody with a design background, uh, building in this way and then releasing it. Um, now I think two things stuck out with me, you know, in like wanting to use the project myself. Uh, one was the token cost. So there's a table that shows like how many tokens you should expect to use each month depending on what mode you're using it in. Yeah, there you go. So your token match budget goes up to $1,000 per month with Opus 4.7. Your conservative budget starts at $40 a month with Haiku 4.5. So, you know, we're using a lot of tokens here, is the, you know, no doubt. And this gets into the architecture, which I thought was also really cool. So if you scroll down, there's like a section on how it works and it shows this like really complicated, um, like sort of logic tree that's going to like try to, to like make sense of data using a lot of different techniques. And it's basically throwing compute at the problem. Um, and so like, like a lot of compute at the problem. Um, and so you have this kind of like big non-deterministic, uh, thing that's like trying to make sense of like the meeting that you just had or, or like, you know, your contact list or something like that. And that, that part has been availed with that precision and recall evaluations. And they actually created their own repo for those evaluations, which is also really cool. The thing is that, that, that, that like front part that uses like, you know, kind of a token matching approach to like categorize your information is like just for exploring and kind of figuring out the workflows that might work for you. So you would be, you know, like typing queries into your agent. You'd be seeing like, is this useful? Do I want to do this again? And if you do want to do it again, and if it's a repeating thing, then you use this minions construct. And minions is kind of like a form of agent that is actually just like a database structure. Maybe it says store procedures that can run procedures over and over again for much less cost. So it uses much fewer tokens. So it's sort of like a cached, like, this works and let's run it again type of job. And then you can further, like, you know, sort of graduate your workflows. You take them to minions and then you take them to, like, queue jobs. Oh yeah, they have a rule here, the routing rule. You were just looking at it. Deterministic means judgment and then subagents. So this is a routing rule that actively figures out, like, which of these What's like the cheapest sort of system that I can run my request with, given my history with the system? So that's really cool. And that reminded me a bit of Elon's algorithm and the, you know, just gradual automation of things that get it to work first, see if it is valuable at all, and then move it incrementally into the automation phases until you have a fully automated system with closed loops. So cool project. Yeah, I thought very interesting. I haven't been interested enough to pick it up and use it myself, but I just like, you know, have enjoyed. And I think that's what a lot of people, if I look at the activity, a lot of people are just like, this is cool, let me play around with it.

Yeah, yeah, yeah. It, it's cool that Gary has shipped this as president of YC. Just like that, that's a cool thing in itself, that technology has made it to where somebody like that could just build something like this for their own personal case and throw it out on the internet for other people to look at and use, potentially innovate on. So, um, I just think that commenting on how AI has democratized creation, even with like super busy executives, uh, you could totally go make your own GBrain. That, that's what this proves. Like, go play around. Like, he's clearly playing around with these personal agents, which I think is really cool. Um, because it's, it's really clear when you talk to somebody that is AI-pilled or not. Uh, and my, my usual thing is like, have they tried to get these personal agents working? And this guy has certainly tried. Yeah, this is cool. So GStack, do you want to talk about GStack? I haven't looked into that that much. We're looking at the GBrain, um, repo right now, but GStack is for coding skills. This has 70,000 stars and 30,000 developers. Using it every day.

Yeah, I guess that's where he started, uh, open sourcing his AI work here. Um, and this one, like, yeah, it says 23 opinionated tools. Uh, it, it's emulating your CEO, your designer, so you're basically kind of like using the whole team to evaluate your work. Um, and you know, this— the CEO role is actually like, does, um, Does what I'm building achieve like product market fit and does it move my business forward? So these are pretty cool. Like, you know, it's a good framework for, I guess, building expertise that can give you feedback as you're building your company as a startup founder. And so he announced this on Twitter, like it got a lot of sort of like interest and discussion around it. You can see there's a lot of people who've kind of checked it out. Yeah.

Um, and you, you could see this like difference between in 2013 when he built Bookface, 772 contributions that year. And then this last year, just this year, I should say, in 2026, 1,237 contributions and counting. And this is really interesting. Uh, I've been building products for 20 years and right now I'm shipping more products than I ever have in the last 60 days. 3 production services, 40+ shipped features. Part-time while running YC full-time. On logical code change, not raw lines of code, which AI inflates, my 2026 run rate is 810x my 2013 pace. 11,417 versus 14 logical lines per day. So in 2026, he's already outproduced his 2013 year by 240x. Um, yeah, not bad at all. I like this. The LOC critics aren't wrong that raw line counts inflate with AI. They are wrong that normalized for inflation, I'm less productive. I'm more productive by a lot. Full methodology, caveats, and reproduction script here. So he uses this. I've never tried this. This could be interesting to give a shot. Um, it reminds me of another AI executive gave me this, which was using AI to create a board to go through the Socratic method with you to sort of get to the root of your beliefs and figure out if an idea is a really good idea or, or not, if you're biased. You know, I think this is an interesting thing, which is looks like you're just getting a whole team to back you via agents.

So yeah, it's definitely a move towards that, like, you know, maybe one-person company modality that Open Claw has popularized. So I guess one question I always have when I see projects like this is like how, how useful are the personas that are created and how were the evals created around them? You know, are the right personas, are they representative? Sort of how do you know that it's working, that it's optimal? Have you tested the framework against other frameworks? I think that the world in this direction is sort of just not very mature. It's very experimental, you know? And so that's what I take it as. It's an experimental approach. Probably if I was going to use this, I would start modifying the personas right away and like, you know, sort of make sure that they work for my use cases. But yeah, it's like this whole, this whole like, let's, let's configure all the agents or, you know, like the people in my company as agents or as like sub-agents is definitely an interesting approach. And then I've also seen, like you talked about, that sort of like let's have a debate or a dialectic between the agents so that it can, you know, now we're maxing tokens, just like having these personas talk to one another, um, and see if that improves the result.

This is kind of cool. He added a /codex for, uh, to inculcate that meta of Codex code review. He just added the skill for it. Um, and towards what you said about the different personas, yeah, I think these are the things to really eval for your case. He's calling them specialist agents, which kind of adapt to the OpenAI SDK, like Agents SDK terminology. These are like specialist agents that help you do a particular domain niche thing, like design or CEOing or whatever. Um, this is interesting too, which is, uh, some standard CLIs for workflows. Benchmarking is an interesting one. Uh, so run the same prompt through Claude, GPT via Codex CLI, and Gemini to compare latency, tokens, cost, and optionally LLM judge quality score.

It's great, kind of like, um, shared learning from Gary to the world. If I, if I could bet on a poly market, if Gary Tan would be using GStack and GBrain one year from now, um, in, in their current form, I'd probably take the no bet. 'Cause I think he'll probably wanna reinvent these, you know, like with the upcoming models that are coming out.

Yeah, I think this is the 2026 meta and by end of 2026, I bet there'll be new stuff around this. This is funny to see Claude, number 1 contributor by far. Makes sense. Cool. Very cool. That, that also is just really awesome because it shows you what's possible. I think that's probably the coolest part about this. Gary is a busy-ass guy, right? Like, he's running Y Combinator as president and CEO, and he's doing a lot, and this is just part-time he shipped this. So for any of the execs or builders that don't have time to play with tools, you have time. If this guy's got time, you got time. So cool. I think that wraps up our AI news segment. Let's move into memes. This is the segment that keeps us sane. So what I'm going to do is actually just share a tab of different memes that I found. And, uh, I found this article with some memes, and then we— I found a few of them that I think are funny, so I'll just show them and we'll add these to the show notes. As well as the video. ChatGPT, can you write original code? No, but you can't either. That's funny. And these are just memes about AI. And in particular, for AI builders, this is a meme around we need to improve our product in a conference room. Somebody says AI, another person says AI, and last person says understand our customers' needs. That person gets thrown out the window. If the answer is customer needs versus talk to AI, clearly you're not going to have a job. This here is called AI. Two older people trying to figure out the computer. You ask it questions and it confidently lies to you. That is so true. Yeah. Yeah. Especially with the whole— remember the whole sycophancy thing last year when they took out 4.0 and everybody was like, no, now it's checking me. It's like making me I was not lying to me, damn it. That was the whole thing. That revealed humanity's bias for ego. That was humanity back there.

Yeah, it's true. This false confidence is kind of the same, gives you the same sort of assurance as the Sikovitsky did, right?

Yeah, it does. ChatGPT calls me by name now. So I guess you can say things are getting pretty serious. This is super funny. Actually, that brings up a whole nother thing. What if I asked ChatGPT to generate an image of itself and Anthropic as dating partners? That— I'm gonna do that. That's an interesting— okay, I'm gonna do that real quick. Create a visual image of you and Claude as dating partners. Oh, I wonder if I'll do it and showcase the differences. We'll see if this goes. Hang on, I'm just gonna share. It's going, it's creating the image. It might stop, but I think I got the animation showing me that this is not gonna stop, so we'll see what happens. Uh, while this goes, let's read the next meme. I'm sure they will use my personal data responsibly. And then a bunch of old rich white dudes, seemingly— I'm not familiar with this meme— laughing their asses off. These are AI companies, by the way. Uh, you've signed away all your rights using, using these tools. Uh, we have to just be mindful of that. Is that right?

Is that Reagan and Bush, uh, like on the left and right? Hang on.

Yeah, I could, uh, Reagan, Bush.

It's like the Republican '80s.

Oh, it is, it is. Yeah. Yes, '84.

By the way, uh, Gemini on my phone is so good at telling me what anything is I take a picture of. Is like, it knows everybody. I can take a picture of a movie, I can take a picture of a flower.

Wow.

Yeah, I have all the— yeah, like I just use the shortcut like you know, what's the— take a picture and tell me what it is. Right? 100% of the time.

Do you wanna see what ChatGPT came up with, with it and Claude as dating partners?

Of course.

It totally took it as we are dating. Claude and ChatGPT are dating.

I thought that's what you meant too.

No, that's not what I meant. I was thinking of it like cuz that meme was kind of saying that ChatGPT and this guy are close. So it's like it dating ChatGPT, but now I'm gonna, okay, now I'm gonna say, no, I mean as separate dating partners for an AI builder to date in a meme funny way. All right, let's see what it gives me for that.

I kind of like that first one.

Like that was pretty cool. This is kind of funny though.

Like, is he saying I have to always have to review your PRs?

I'll send 10 memes in a row. Claude, I'm not rushing, but here's a thorough take. Love language, words of affirmation. We were just talking about the sycophancy on ChatGPT side. Yeah, I think that's true. Claude, thoughtful insights, remembering details. I definitely think that's true. ChatGPT, hype and encouragement. Yep, for sure. Strengths, brainstorming anything, explaining complex ideas creatively. Or simply creative ideas, always up for fun, quick on my feet. Strengths for Claude: deep analysis, well-structured thinking, cautious and careful, great at complex solving problems. Typical dates for ChatGPT: deep talks, trying new things, spontaneous plans, lots of laughs. Wait, let's research this real quick. For Claude: cozy nights in, deep conversations, planning the future, reading together. Let's make a plan and stick to it. So funny. Wait.

Do you see GPT's coffee cup?

Yeah, that's so great. Overthink out loud is what the coffee cup says.

That's pretty good.

Like, that's—

I think I'm surprised that it's so introspective.

Two different AIs, one amazing connection. Thanks for bringing us together. All right, now let's see what this has. No, so this— I prompted it saying, no, I mean as separate dating partners for an AI builder to date in a meme funny way. ChatGPT, though, shit. The hype boyfriend: Babe, I wrote a poem, a business plan, and a roast for your ex. Claude, the thoughtful boyfriend: Before we build anything, how are you feeling? ChatGPT, I could do anything. Claude, let's figure this out together. I think that's actually— that is everything there. I think that says a lot about The two, at least in the way I've used them.

So either way you're winning, just don't ask them to fight.

Look at this guy. He has a pizza box, pizza and props. Claude's got like a Bible.

That's really funny. So what was it? Ethics of Everything? Is that what it says? The Ethics of Everything. Philosophy of— morality.

Yeah, yeah, yeah, that kind of jives with all the PR. Funny. We'll, uh, we'll include these memes into our show notes.

More tokens, more problems. The coffee cups are just killer.

They're pretty great. They're pretty good. Um, yo, check this out. So the Sam Altman— next meme— the Sam Altman and Elon Musk trial. Had been going on, right? And near me in the Alameda courthouse.

They just reached a verdict, by the way.

Yeah. Yeah.

The jury dismissed all charges. Yeah.

Yes. Yeah. Yeah. So we— yes. And then I saw this meme, which was interesting, but when was that released? Was that today?

Like when we started talking, like 30 minutes ago.

Oh. Oh, my gosh. That was just 30 minutes ago. I saw something about it. Okay. Yeah. New York Times OpenAI lawsuit live updates: jury rejects Elon Musk's claim in a unanimous verdict. Elon Musk loses court battle against Sam Altman and OpenAI after 3 weeks. Okay, well, this is interesting then, this meme. Uh, much of Altman's time on the stand concerned his relationship with Musk, who is pretty much the same behind closed doors as he is in his public persona. LMAO, Altman says there was a meeting at Tesla during the evening about folding OpenAI into Tesla for AI research, and then a long period of time with Elon showing us memes on his phone. The court reporter has to ask him to repeat memes on his phone loudly. This is from somebody, uh, giving a play-by-play of the actual trial. And here's, uh, what, what an artist's depiction of the, uh, trial look like? Sam Altman, slightly bigger head than body. Yep, fair. Elon Musk with a cat meme. I think that's what this is supposed to depict. Anyway, what do you think of the OpenAI, uh, Elon Musk trial and the verdict?

That's a pretty funny meme, I think. That, um, yeah, like I, yeah, I don't know how deeply Elon goes in some of these strategic conversations when he's, when he was like trying to wrangle control of OpenAI and he's like, let's turn it into a for-profit and I want to run it. Like, yeah, I don't know that he had a lot of like depth of thought other than like, I don't trust these guys, I want to run it. Like, and that seems to be his modality, you know, with all of his, like, he doesn't have like really peers in his company, right? Collaborators. He really is like, he works well as like a sort of a bottom down, you know, bottom up, like type, like dictator, like hierarchical.

Yeah.

Yeah.

Control.

And that's, I mean, he succeeded that way. So he's, you know, there's not, there's not any pretending that he's like trying to work alongside anybody or get that. You know, or, you know, somebody that doesn't get along with him has an unreasonable point of view. It's just like, no, you're just on Elon's show and he's going to run the show if you're here.

Yeah, it's definitely like main character sort of energy is what he exudes. Yeah. So I wouldn't— I think he is a deeply thoughtful guy from my reading of him and his book and stuff. Like, he certainly wants us— he wants more for humanity. From what I could tell. And you have to be deeply thoughtful if you're trying to build rockets and you succeed with that, right? Like, but it is funny that on the flip side, he is— they were just talking in the trial for a while about him showing Sam Altman memes back in the day. Like, that's like, I could totally see this happening. And then I could totally see him just thinking, yeah, these guys blow, I'm gonna try this instead. And then he got a little upset, didn't go his way, and then now we're here. So yeah, such is life. We're all, we're all humans at the end of the day. Yeah, just want to show my friends my memes, and then if they're trying to take over the world and I think they're doing wrong things, I guess I'll do things about it in my rich person ways.

Yeah, totally. I think I heard, um, it was Ben Thompson talking about the trial last night, and like the it takes such a winding course, like, you know, and there's a lot of things you could sort of look on the outside looking in and scrutinize about things that OpenAI did. But, you know, for the most part, they made pretty rational, reasonable decisions and defensible decisions the whole way through. And I think that, like, that, that sense of them being a rational actor did not get destroyed during the case. And I think that's probably why, you know, there wasn't any kind of repercussions. Um, they just acted in the best interest of their business and, um, in the most rational way that they had available to them.

It's cool that this has wrapped up. I only slightly regret not going to see the clown show in person, but once in a career opportunity. Yeah, that was— man, I could have caught— I could have caught some memes in person. All right, cool. Last couple things. We're going to move into the last segment of our show for our weekly AI Clarity, uh, which is What's the Meta? In this segment, we talk through what does it mean if you're building, if you're an AI builder building on the front lines just like we are. We consume the news, we look at new meta updates and how people are building, and we bring it to you weekly, all the new updates on what's the meta. So this week we have a couple items in What's the Meta. A staff engineer's honest AI report. Let's go ahead and up that. How I use LLMs as a staff engineer in 2026. This is by Sean Godek, or G, if I'm not pronouncing that right. Sean says, a bit over a year ago, I wrote how I use LLMs as a staff engineer. That article went pretty viral. And this is his update for 2026. Um, for last year in 2025, he says he used for AI as of last year, smart autocomplete with GitHub Copilot, short tactical changes in areas I don't know well, always reviewed by a subject matter expert, writing lots of use-once-and-throw-away research code, asking lots of questions to learn about new topics like the Unity game engine. Last resort bug fixes just in case it can figure out things immediately, big picture proofreading for long-form English communication. And here are some tasks he explicitly did not use for AI last year: writing whole PRs for me in areas I'm familiar with, writing ADRs or architectural decision records or other technical communications, research in large code bases and finding out how things are done. So February 2025, that was, uh, so last year, and eons have happened in between. And he's basically asking what's changed since then and now. And he says agents are good now. I love that this was a headline, just bold heading, agents are good now. And we talked about this earlier on the pod, uh, Jonathan was saying, you know, in September he was testing things and could we just let an agent just drive everything and we could build apps? And it was like, clearly no. But as of December, when Opus dropped, and then all the big acceleration of LLM model, uh, I think capacity for coding agents, I should say, really improved. Um, agents are good now. That's a great way to put it. So kind of, kind of moving ahead, I'll summarize what he uses AI for now, and then me and Jonathan will give our own takes. Overall, in 2026, what's Sean using AI for now as a staff engineer? Writing or drafting, depending on complexity, every code change I make. Investigating and fixing bugs, either autonomously for most bugs or with my close involvement for trickier ones. Research in large codebases since current agents are now good enough to give the right answer almost all the time, and when they're wrong, it's clear from reading the explanation that they missed something. Manual testing and local machine setup or troubleshooting. And I still use AI for asking lots of questions to learn about topics and for proofreading. What Sean does not use AI for in 2026, still writing any kind of public communication for me, pull requests, descriptions, architectural decision records, messages, with the exception of trivial 2-line PRs. Writing code that I don't carefully review, testing any kind of UI. In my view, the current core AI skill is shifting as much work onto AI agents as possible without going too far. Many people are underutilizing agents, not allowing them to investigate bugs or test their changes, or not throwing enough simple tasks at them. Other people, Sean says, are overutilizing them, using them to write messages that ought to be handwritten, or trusting them to make sweeping changes that need careful human review. Since the 2025 post, Sean says the balance has tilted more towards the agents, but finding the balance remains as tricky as ever. All right, so for What's the Meta, since we're AI builders ourselves, thoughts on this take, Jonathan?

Yeah, it seems very pragmatic and it's cool to see it year over year. Hopefully we'll get like a 2027 version from him too. And I guess the, the most interesting part that relates to my own work is that the stuff he doesn't use AI for, I use AI for some of that stuff, and, um, it's, uh, it's questionable.

So, you know, questionable which way, that he's not using it or that we are?

You— well, yeah, like, say, um, I— like, say, as of like late last year, I had a rule I would review every line of code that, that the AI wrote. Um, I've abandoned that rule because I can write so much code so quickly and sort of like, you know, there's a higher level of like, are the tests doing the right thing and are they passing? And that's more the kind of heuristic that I'm looking at as I'm generating code now. And I don't— it's sort of like, I don't feel like my code quality is as high as it was before, but I can write a whole lot more and I can usually get it to the quality if I invest enough time. So there's kind of like a trade-off, you know, it's a little bit like our first conversation today. And then the writing any kind of public communication for PR, PR descriptions, ADRs, you know, I think we do a lot of that in general. Like our whole linear ticket backlog is like, you know, written by, written by AI. And for me, like when I write a ticket or a project, I might redraft it like, you know, a dozen times to make sure that I like the way that it sounds. But I don't want to actually write it myself. So I feel like that's leveraging the tools and figuring out like what my taste profiles are and stuff like that. I can be more prolific that way personally. So yeah, those are two pretty big differences in sort of where our independent judgments as, you know, as engineers using the tool are.

What about the testing of UIs? So that's like writing and then— Yeah.

Yeah, but the testing of UIs, I found like a kind of love-hate relationship with this because I mean, uh, I think Chrome DevTools and Claude Code was like a step function and like, oh, we can test UIs now. And I've got a lot of productive testing done that way. Um, it can only get us so far. And I also feel like my, probably my front end, uh, QA review skills have atrophied a bit as I've relied on the, the, on Claude Code. And then Codex recently came out with their vision tools and they have desktop browser integration. Actually activated at the same time as I did the mobile phone this last week. And, you know, I was just really impressed at what it could test and how it could sort of run an Aistrow test against UI. Does it catch everything? No. I mean, but like it catches a whole lot of stuff and we make sure that the system is basically working for its prescribed purposes on staging before we push out to production. That's really helpful. So, yeah, I think I'm— the first two I use like probably prolifically. And the third one, I wish I used it more, but it's just kind of slow. You know, the UI testing is still pretty slow.

Yeah. Yeah. So I guess I'll comment. You got Jonathan, our CTO's take. I definitely use this to write public communication, like, for sure. And there's— I— so his take on I'll use this, I'll use this to write things, but then I'll, for very sensitive communications, he deems these ones sensitive, like PR descriptions, ADRs, Slack messages, uh, he chooses to write them himself. This is also a staff engineer at like a big company sort of situation, like kind of different, you know? Um, my take on this is that I have to draft so many client updates, uh, cause we're operating as AI builders and for deployed engineers. It would be impossible to— I would be spending way more time drafting comms and getting tickets updated and stuff like that. Linear MCP with Cloud Code with Google Workspace MCP, so you could draft emails. Um, I write— if I have to give a Slack update to a client or to Jonathan, uh, or contractors on the team, I'll have it draft me a Slack message and store it in my Obsidian, my personal That's usually my meta. And then I'll copy paste that Slack formatted thing into Slack and then I'll tweak some stuff. But like Jonathan says, I'm usually drafting these comms with the AI. Um, I'd say for my other stuff, like social content, definitely drive that, the draft generation with AI a lot. And I've seen some misses where it's gotten through a quality review for me. But my trade-off is like, if I could do this for our content engine, and it keeps me consistent, and if everybody's doing it, then it's just more about, okay, at some point you figure out the evals and the harness for every content case you have, and then it gets better. And then you scale and leverage yourself infinitely, right? And ultimately, we're a startup. So scaling and leveraging infinitely. Sounds pretty fucking great. So I'm gonna keep writing with AI, and then curating with my own human taste. And then evolving my own harnesses around that. Writing code I don't carefully review. So yeah, definitely have abandoned any sort of pretenses of reviewing code myself. I review a lot of English. And I think it does come down to just those things like, Um, can I give the agent the right context? So we try to define the customer requests we get in Linear and have that very— we, we will, you know, go from, had a meeting, we recorded the meeting transcript, we try to AI prompt Claude Code to take the meeting transcript to create customer requests in Linear to evidence against what a customer actually said to, you know, tie out the meeting, uh, the the customer request to first-party evidence. Then from that, once we feel pretty good about the functional requirements, we use Cloud Code with the context of our codebases to then enrich it with technical context, right? Then at that point, it's sort of like a baked piece of bread. And then you can go into, let's take that context, which has the functional and business side and some of the technical high-level side, and then let's massage this into a spec. For the pull request. In our pull request lifecycle, our input docs are implementation plan, which has sort of the checklist of all the files and changes to make, the research, which informs that, research.md, which researches the existing state of the codebase at this point in time. And then a test strategy.md, which has the exit criteria is the way I see it. Like, if we gave this a goal or a Ralph loop, and we're like, really, just get to here until you could do these things. And here's the acceptance criteria, exit criteria, uh, and tests. Like, we gotta evolve our tests, unit integration tests or end-to-end tests. Um, so I do a lot of reading on those documents, but I don't do a lot of reading on the code that is produced. And then I do a lot of like, uh, because I'm dealing with like mobile app stuff right now, a lot of just end-to-end testing on devices, which is endlessly annoying. So the way I've sort of— my meta there is, um, I'll run multiple iOS simulators in parallel towards a chunked out PR for a particular feature. Then I'll have it look at the exit criteria and say, test the exit criteria once it's implemented the implementation plan. MCP iOS simulator has gotten decent. The screenshotting sometimes fills up the context window, so I have to be mindful of that. You have to massage the context. Or even start it off on the right page because they haven't like optimized that MCP is my experience. So testing any kind of UI, I definitely use agents, but ultimately what goes to production is I've like verified this on my own device because when I haven't done that, it hasn't gone well. Can't trust it. YOLOing it to the gods, the AI UI testing gods yet. So yeah, but maybe, you know, we've kind of made it better.

Yeah, totally. I think we've kind of made a decision as a company to be AI native, to be on the front edge, to try to use this for everything that we can. And so like, you know, there's, there's an aspect of like where whether it's code reviews or right, generating writing or doing UI testing that We're like, you know, let's learn how to do this and see, push the tools as far as they can before we decide not to do it. Right. And I think in both of these directions, we haven't decided not to start, not, not to use agents for UI testing. We haven't decided not to use agents for— they're generally serving us pretty well, although there's recognized shortcomings.

Yeah. Yeah. And It's sort of just having the right principles to drive the decision-making. Where we're like, at least this is my sense, where we have enough agency as builders where we're like, I have a good sense at this point in my career, and you're at this— I'm at 10+ years in my career, you're in like 20, 25+ years in your career. And so we both as founders have a good felt sense of like, I'm going down too far down this rabbit hole, we should cut come back up kind of thing, because I actually got to ship some shit to production to get it working. So there's a— I think there's a good seesaw of balance we both have where we reflect, even on these pods, we'll reflect, or in our weekly team meetings, we'll reflect on like, hey, like, what's working for you AI-wise, or I'm running into these constraints. Um, but sort of my rule of thumb is if it feels like this is taking way too long at a certain point, like it feels like this is I'm like an hour in and this doesn't make any sense. I'll come out for air and be like, all right, well, I actually got to get some shit done. Like, let's go back to proven patterns, right? But other times I do encourage us to explore because if we're going to win against incumbents, it will be about our exploration and being on the edge of using these tools for the arbitrage they provide for us to go faster than the competition. That's how I view it. So, so far, we haven't had to, because me and you are, we trust each other. And we're both high agency founders and builders. We haven't really needed to be like, alright, right, like, here's like an inculcated principle for how not to go down this rabbit hole. But I could see that happening as we, you know, build out the team going forward. So yeah. Yeah, that makes a lot of sense. Any other thoughts on Sean's article here about what he does and doesn't use AI for?

No, it seemed very pragmatic, well written. I enjoyed reading through it. So yeah, good stuff.

Yeah, sweet. Let's move into one more thing. I think we'll have time for that. Vercel Labs, they just released this AI They just released an agent-native programming language. It's called Zero. Um, this is the GitHub repo, already has 2.2 thousand stars, released last week. It is an experiment in building an agent-first programming language. The project is exploring what changes when agents are primary users from day one. A language that can be learned on the fly to tooling that exposes structured facts for debugging and repair, and a standard library broad enough that most programs do not start with the dependency search. Zero is pre-1 and intentionally unstable. So breaking changes forthcoming for sure if you're using this. But I think this is, this is really cool. Here's like a marketing post by MarkTechPost. So, uh, this is, this is a cool image. Where the CLI emits structured JSON by default. So this is zero-check --json, and it'll actually throw out particular structure for the agent to check out, like code, a repair. And so what they say here is that each diagnostic carries a stable code parameter, a human-readable message parameter. Parameter and a repair object with a typed repair ID. Humans can read the message, agents read code and repair. And it's the same CLI command serving both of those personas. Yeah, I thought that was, I thought that was pretty cool. I don't know where this is gonna go. But it makes sense that somebody would try this. And I was surprised. It doesn't surprise me that it's Vercel. But when I saw it, I was like, oh, wow, this is cool. Why would they try to make this? And then From a business standpoint, it totally makes sense if other people adopted this and they own sort of the deployment and hosting story. That'd be pretty dope. And what I imagine is if an application adopted this language and then deployed it, hosted on Vercel, you would probably get better, less erroneous build pipeline over time. That's what I imagine would come out of this. What do you think of this zero programming language for agents, Jonathan?

Um, yeah, it's a really interesting idea. Uh, it definitely serves a purpose, I think, just in terms of efficiency. Um, you know, like, we've been writing code for years at the human level, but, you know, like, when we say compile code in our browser, you know, JavaScript gets minified, and, you know, like, we can't read that necessarily nor debug it. But machines can. Um, and then the format, it was kind of interesting because the, uh, the harness that I'm having Codex build right now, this is what the code looks like because it's, it's building a harness for like an agent description to, to another agent, right? Um, and so it's kind of curious that like, yeah, this, this is kind of structured JSON-looking, uh, blob is, is like the best way that, that agents have to interpret it. It matches my own experience about how my agents are talking to each other. We'll see where it goes, you know. I don't know that it even matters. Like, I don't even know that humans need to define this. Like, I think that agents might just be able to figure out how to talk to themselves more efficiently without us. Yeah.

That's funny. Yeah. Hang on. I'm gonna ask ChatGPT to create agent programs talking to each other. Right. Agent programs talking to each other. What data is talked?

Yeah, that's kind of like—

have you heard— go ahead.

Have you heard that like economic agents will just create their own like crypto chains to transact with one another?

No, what?

That's interesting. Yeah, that's, that's one of the theories about like, you know, how, uh, agentic agents become economic is that they can actually just spin up their own liquidity pools with the crypto token and, you know, like transact using those. They don't need to use any of the existing chains. Um, so this is why I say like, you know, it's, it's cool that Vercel wanted to define this, but it's probably something that two capable agents can do on a bespoke basis with one another on the fly, uh, because they have all the information they need to create an optimal communication language.

Interesting. Yeah, that makes sense. That makes sense. Hang on, check this out. Looks like they are talking to JSON. I didn't tell it to. Agent to agent protocol summit. Let's talk data. Hey Planner, what's the plan? Goals, constraint, budgets. Bunch of JSON. I found these options. Oh, this is kind of cute actually. I like this. Yeah, this is actually a great way to show somebody how this shit works. Like, there's an— there, there actually should be like— I guess this search agent is the orchestration agent because it's then going to the plan. Okay. And then plan finds options, evaluator figures out best options for the planning agent. Okay. And then passes off to a booking agent, logging agent, just logs the whole time.

The query is find the cheap flight to Paris.

Yeah. Wait, do you see that somewhere?

Oh, it's on their, their little table.

Oh, here. Oh, okay. I didn't need this. Find me a cheap flight to Paris. User profile loyalty none. See preferences. Aisle USD currency. Time. Tools available. Oh, that's cool. This is a great way to explain this to people. Actually, I'm gonna say this. This is, this is really good. Cool. That's sick. Pod notes here.

Yeah, it's, that's very tweetable.

Very tweetable. Yeah, I'm going to tweet this later. Um, so yeah, the problem that they're aiming to solve is that most compilers emit unstructured text errors written for human readers, so agents have to parse that. That's fragile because error formats could change, APIs change, messages are human-oriented. So this is, this is about emitting structured JSON diagnostics by default, and that's really cool that We had come to, in our stacked agent harness architecture for Clarity, to allow people to build their own closed-loop autonomous agents towards any particular closed loop and then incubate those agents and improve them. That's what we're building with our architecture. You kind of came to the JSON diagnostics language-ish as an independent finding. So that's cool. Makes a lot of sense. Yeah, it makes a lot of sense.

Like, does that— that your harness specification should be in a language like this? Like, that just makes sense.

Cool. Okay, I think that's probably good for today because we've been going for so long. I think we'll go ahead and end it here with this agent-to-agent protocol summit image, which is a great how the fuck do agents work. Probably just send this to somebody and show them what is the user query in the middle, and then all the agents having coffee, sharing bread, and doing stuff. And then, then you get your answer and you get a flight booked. This is exactly what happens. This is what happens at OpenAI. They have these little cute agent bots just in the coffee rooms taking queries. It's exactly what happens. So please share with other AI builders so they are in the know. None of it is magic. This is how it looks. Cool guys, thanks a lot Jonathan, and thanks for listening to the pod. If you have AI memes, if you have AI news, if you have new meta for AI builders that you want us to showcase on the pod and give our own takes, please submit those to podcast@epistemicme.ai. I'll read every single entry, or throw them into comments on YouTube or anywhere else you can find us on Spotify or Apple Podcasts or your favorite podcasting platform. We will see you next week and keep building. Take care.

Bye-bye.