⚡ Inside GitHub’s AI Revolution: Jared Palmer Reveals Agent HQ & The Future of Coding Agents

I'm only on I'm only full disclaimer.

Don't let that. It's from from the interim. Okay.

I spent the last sort of two years or so building v zero at Vercel and AI SDK. And then the summer took time off and now joined GitHub. And today we launched AgentHQ among other things here at Universe.

And, yeah, it's gonna be the home, we hope, of not only agents, but also developers. And this seems like a the gravity well of this new collaboration space that we're trying to build. Yeah.

It's some above average. Yeah. It's under 8,000,000 developers.

Pretty. It's just the scale is immense. Right?

And vZero was focused on not only one language, but one framework. Right. And a specific problem space with a built in renderer and, you know, you know, for those who are not aware of vZero, it's like a Bolt or Lovable, but it's built by Vercel.

And it's focused on building Next. Js apps, specifically Next. Js apps.

That constraint was rather liberating for the team at the time. And it lets us really like laser focus. Does.

Your world at it. I hope. Thank you.

I hope so. And obviously at GitHub, you know, we're the home of all languages and, you know, frameworks and developers. And so the scope is broadened.

different parts of it in my brain by going back, you know, like, so maybe we'll have to figure out how retrieval never worked. By the way, interesting how memory works for agents. Totally.

That's why I brought it up. This is kind of interesting as you Yeah. Sometimes you discover new paths.

Right? Yeah. Anyway, the story goes like this.

So when ChatGPC first came out, obviously it was incredible. Right? Like world changing immediately, faster growing product ever.

I looked back at like the timeline and dates, and we were very early, like in when I was at Purcell, jumping into AI stuff. But the journey kind of went like this. So at the time, I actually, there was no AI division.

There was no AI group. I was actually the director of engineering for all of Purcell frameworks, and I was helping Next. Spelled spelled right.

Yeah. So Next. Js, Svelte, TurboRepo, TurboPack, Webpack, and all internal dev tools at Vercel.

And I was helping the Next. Js team dog food and test the initial implementation of server actions. And instead of building a to do app, Guillermo, the CEO of our cell was like, why don't you build like a playground?

And I was like, okay, cool. So that led to the AI playground, which is now just part of AI SDK. We'll get there in a second.

Which by the way, iconic for like side by side, but also the Right. So Nat. I've seen so so I'll I'll It's Nat.

Dev. Nat. Dev.

So so g g told me that Nat Day of I got I got a DM. It was like, remember because I was at a bachelor party and Gijermo totally online, sends me a note like, Nat. Dev is launching on Monday.

You have to ship. And I've been working on it previously. And so he was like Same idea.

Yeah. So he got wind of it, I guess. So I I definitely had to like jump into motion.

And I didn't think we we didn't even ship chat first. Gisermo sends me this this DM over the weekend. I'm at a master party, and he's like, nat dot dev, this the side by side.

He sends me the link, and I play with it, I like, ciao. Okay. So I was bringing the gear, ship the AI Playground.

What was cool about the AI Playground was it forced me to go through every single model provider's API docs, and figure out their their quirks of their Nuance streaming. Because at the time, it wasn't like everybody used OpenAI. Yeah.

It was like, oh, little quirks. Some of them kinda were compatible. So that was my first foray.

And then launched AI Playground, that shot to the top of Hacker News. And I remember we didn't, I didn't even implement chat because that chat wasn't actually like a fake, like wasn't as important. It was just like completion of lesions.

So eventually we we factored the chat and that that project, out of that came AI SDK because I had already looked at all the model providers and all the the combinations. I was like, okay, here's that chunk of streaming code you need. And then AI SDK found that sort of niche of like, how do we focus on the part that we're gonna be good at, which is like that UI aspect of it, but then also knock it in your way.

So that we shipped AF Playground, then AI SDK launched. And then, you know, we're always about demos and having great starter templates at Vercel. And I and I remember writing Guislain.

I was like, you know what'd be cool? This guy, Chad CN. Oh, man.

He seems like amazing. And his UI library is doing great. Why don't we team up and ship a ChatGPT clone open source?

And we did. We shipped this this this awesome template, which is now called Chat SDK, but it's great. And what that did though at Vercel was it set us up for like rapid experimentation, because we had this really good, like pretty full featured ChatGPT ready to rock with all the latest features.

Yeah. So when it came to like rapid prototyping that summer, now we're summer twenty three, it was so great. It was it was like liberating.

So I remember at that point, I had gained some momentum internally and pivoted almost entirely to AI. And I had Shu Ding, who you're friends with, and Max Leiter, and Shad CN now were cooking. And I think at that point, code execution had just come out.

I think that's my timeline. Yeah. The cool sandbox interpreter.

Inter code interpreter. That's what they call it at a time. And I had a very as soon as I saw this, I had a very ambitious idea and proposal to present to Guillermo, which was like, what if there was some and mind you, tool calls don't exist.

The context window is 4,000 Wow. Tokens. You know, like there's not much here.

What if we had this thing where like you could prompt, and sometimes it would do code interpretation, and then maybe we could sort of sometimes it would decode up interpretation, but then other times it would choose to render like UI, or then it would render sometimes like a document or Same line in the chat. Yeah. And like, would just have different sort of Generative UI.

Yeah. You could pipe them together. So like the output of one could pipe into so if we did code interpretation, we coarse it to always emit like tabular data.

Yeah. Maybe we could pass that to another prompt that would like a UI. And just like some idea there, it's got crazy.

But if it sounds like these are just tool calls, that's exactly what these Yeah. It became pretty obvious that like the so I came back to Jeremiah and we at the time, we just had a secure we had a sort of security debate, like, should we code interpretation with the ability to fetch data? Like giving it internet access was like kind of now they're like, Fine, whatever, whatever, do whatever you want, wow, wow, Wes.

But at the time it was like a little scary. So we kind of said, Okay, no to the code interpretation, but this UI thing, it's pretty neat. And so that, this like prompt to UI, that was like the moment of v zero, but the models were not very good.

Right? Or relative to the way they are now. It before GPT-four era?

This is just into the GPT four era, and now we're probably at a 16,000 token context window. So you can't really do chat. So we had to kind of invent this kind of new paradigm of, like, fake it with completion.

But that forced us to do sort of the click the well, the the initial v zero, which launched I think in September 23, it would look more mid journey. In fact, if you go back to the original tweet, it was like mid journey for React because it was all very visual. And you could click on different components and elements and re prompt.

But it was, again, we're we're kinda hacking this because we didn't have chat and we didn't have tool calls. And then fast forward again, you know, that launches. And then probably like nine months later, it took us like nine months to gets like a million ARR, this little team.

But then the models progressed. And from, you know, GPT-four, GPT-four 32 k, the big big boy. We never really got GPT-four turbo working.

I don't know why. It should never happened. And then switched to other other Frontier models and and then started doing our own models and stuff like that.

But fast forward another ten months or nine months or so, and then we we rebase towards chat. And now the models finally could do chat, and the artifact pattern had evolved. So it was time to rewrite.

When we launched v zero, the chat version or the new v zero, whatever you call it's like 14, another million MRR, fourteen days, another million MRR. It was like a rocket ship after And that just proceeded, and we just kept cooking. And so that's been the journey.

We just kept perfecting. And what was really liberating for us was actually the focus on just one stack or one framework. When everybody else was trying to do general purpose coding agent, we were like, no, we're just gonna focus on Next.

Js, Frontend, and ChadCN, and that was really that allowed the team to focus. So that's the that's the story arc. I mean, to be fair, like, bot lovable, like, because Next.

Js is so dominant, basically everyone has to be good at Next. Js. Right.

But being focused on, like, even right down to the UI library and component stuff, like, that that actually helps a lot. We also started working with all the Frontier Model Labs to help, because it was in our Vercel's best interest to have them be created at next.js.

Yeah. And also because of the post trained models, and you can read about this on the Vercel blog, like we can the post training like, harness that we created, we were started sharing and stuff with other model labs and stuff like that, and we had all our data in a very hygienic state to work with them. So Did you ever debate internally?

kinda We went back and forth, and I think yeah. You know, we launched we went back and forth on all this. I think at the end of the day, there's pros and cons.

Yeah. One of the benefits of having your own branded model or synthetic or composite is that you can stitch these things together. Yeah.

There's like higher levels of And now it's a little different with this AgenTic flow, but even looking at what you guys launched recently with ZZRX3 rep, right? So search, is it gonna be a different model than what generation, you know, the Genesis, but like, because it's a shorter. But like search and Genesis are two different mode, like entire subsystems, right?

So you can have search evals that are gonna be totally different. And so how do you, so where we ended up now, or are we? Where we've probably ended up now is like, for a long time we didn't have model selector, and then we had our own models, which were composites, which we talked about.

And then like, and that would allow us to, you know, mix match. And I think that's probably if what it's also nice because you get as a this is like the product cap. You get to brand it.

Yeah. Right? And you can decouple it from the launch of the Frontier Lab.

Yes. How do you guys how does Cognition even bill for it? Like A few years.

Cranky is a Compute Out. Synthetic unit. Right?

Yeah. So it gets a little wonky. Yeah.

We can go on for pricing this this stuff. It's it gets challenging. But the nice thing about having the, like, brand name model is that, like, you get to co launch with the provider and they'll hype you up.

well, we've switching gears to GitHub. Like, we are all about model choice now. Like making sure that and what's cool is that, like, it's we also have Copilot, which is our harness, and Copilot CLI, but we also have third party harnesses like Cloud Code and Codex and Cognition now like in H and HQ.

So you kind of get the best of both worlds. Yeah. And like, I think that's gonna be awesome and ultimately what people want.

Yeah.

no. So my initial definition of agent for AIS, because like I actually fucked. I was dying on this hill.

Because AISDK, everyone else is an agent framework. And I think maybe they actually went to like this. I don't know what it says on the front page now, but like who knows?

But like like six sixty. An agent is, you know, an agent is orchestrating, you know, an API request with a queue and a for loop. Okay.

But a coding agent now has meant so much more. There's like these, you know, coding agent SDKs, and you've got sandboxing and file systems and tool calls. And I do think that is a uniquely, I'll call that agent world, and I'm in charge of that coding agents here.

And, I think that that seems to be where things are going. And even I believe the the Claude Excel agent is basically I was talking to to Mike Rieger backstage. Like, I think it's like related to Claude code.

I think that AgentHQ and GitHub itself needs to, like, co evolve. And you know, one of the things that Microsoft has done really well is by putting things that are alike closer together. And so you think about the new core AI organization.

Yeah. You've got Visual Studio. Okay.

Visual Studio Code, GitHub, and parts of Azure all in one. And obviously, the GitHub team and the Versus Code team have been working closely together for a long time, but now we're really close together. And I think for me, one of the cooler things that AgentHQ can sort of offer is this seamlessness, the fluidity with your workflow.

Right? So if you saw in the demo today, we saw a demonstration of you use AgentHQ, you fire off a task, and it creates a PR, but you can also open that PR up in Versus Code in one click. And that's that's awesome.

And I think my the the vision for for for GitHub as it evolves is to look at those touch points where AI can be sprinkled in, you know, Saltpaste style, into the native workflow, whether you're assigning an issue or maybe some new stuff that I think we should like focus on. Maybe it could be like, how do we resolve a merge conflict? Oh my God.

Right? Like how do we maybe pop open an action or like get in right? And so as like I think it's probably merged, like this is my definition of AGI.

Totally. But like you get that error on an action and you're like, we've all been in that sort of flow where like actions kind of don't work locally, but is that tool act? If you ever tried it, you're like, what the hell?

Don't have that set up on my machine. I haven't done this in a while. So you're like pushing up and you're you're this like, okay.

What if we could just put like, oh, you know, comment or kick off a task, dev solve this or for you, or throw things there. I think what I'm trying to describe is this like, this workflow where it's just like seamless and fluid, and you can stay in a flow state across, whether you're across, like across all devices, mobile, web, on github.com, or in your local editor.

And I think that's where, you know, my focus is gonna be in the next Yeah. Six six months or so. Yeah.

Yeah.

Yeah. I think they were so far ahead almost. Like it was but now there's sand this like sandboxes, there's so many of these days.

Right? So I think PowerFlare just launched theirs. Yeah.

There's Daytona. They're here. Vercel.

Modal, which I think Lovable uses. I I'm not yet. I don't know.

Yeah. I mean, you you probably have your own. I don't know.

What do you what do you guys use? Just some Kubernetes pods. Okay.

You guys are rolling it yourself. Very cool. I think that's maybe the runtime, But but there's there's work and discussion about what that runtime should be even internally at Microsoft.

And we've got a couple different competing things. So we'll figure it out in the next, you know, in the next cycle here. But there's a great point.

Like there there's a lot of cool stuff that is in a dev container. Already got Versus Code loaded. You've got a file system.

You've got a sandbox. You've got the security protocol. Yeah.

But also, like, wired into GitHub Enterprise.

Why is it so hard? Like, why have we solved it? I don't know.

I think it's hard because you can't predict what's in the what's in the repo. Right? So it's like and Yeah.

You don't know when they bundled FFmpeg. Just don't know. Right.

It's nice when, like, if it's just Next. Js, you just run PMPM install. Correct.

Correct. You you can like do special like there's obviously cons through constraints, you can make optimizations. And I think the general purpose container is just like challenging.

That being said though, I think there's probably some work to do on, you know, auto detection and preempting and stuff like to be done there. But it's just a bigger it's a broader problem space. Right?

Yeah. And yeah. So so fun fact.

and then also, like, what are your defaults? They're not exactly the same, which would be better to, like, just having even the same, like, preference stack Yeah. Of defaults Yeah.

Is the right Yeah. Yeah. Yeah.

Would be great. Because then we can move the whole ecosystem together Alright. From, like, the PMPM or bun.

Right?

no. The ZRED one. That one.

Okay. And then the the payment that was it Stripe or Coinbase? Stripe.

Stripe? Yeah. That's very cool.

We've had them on the pod. Okay. Yeah.

That's very cool. It'd be interesting to see if that takes off. I mean, it's Stripe.

It's so Yeah. Yeah. But it's always to be adopted by like the clients.

Right? And Yeah. And and I think that's that's fascinating.

The MCP is huge. It seems like it is the way that a lot of the especially when it comes to like digital transformation or some of our enterprise customers, it's where they're gonna able to be at or where they are able to add context. In addition to that, we also have custom agents that we announced today too.

So like you can work with prompts and stuff within your within your agent HQ and and customize these agents for different tasks, and and those can have MCPs and such. And I think that should be really powerful from, like, a a platform perspective. Yeah.

It gets me excited. That's what, like, I think is, like, shipping now and and the next but we're always on the lookout for, like, the next thing. And I don't know.

What what's on your what's what's top of mind for you? For, like, standards or Standard.

computationally intensive, It's my

just give it a few more turns of the Yeah. It's killing.

them good? Making them good? Yeah.

You don't like cloud cove? No. No.

I I don't. It's more just like, you know, the the devil's in the details. Like Yeah.

Going from 90% going to like hill climbing, it gets steeper. Yeah. Yeah.

In my opinion. It gets steeper. And so going from 90% success to 95 to 98 to 99 to nines of success.

Yeah. I mean, we're really hard. Paying Merkur a lot of money for expert programmers of open source maintainers.

So like, you know Then you realize along the way, maybe the users aren't that good at it. But like yeah. No.

But I I just think there's a lot of work to do to finish the swing. And there's a big difference between ninety eight percent and ninety nine percent correct. And that's like noticeable.

And this used to hit, you know, if you're working on AI product, you probably don't realize how you've probably seen this. Like most people are blind, like living in la la land about how poor quality their AI product lately is. Unless they're really measuring like a number of error free sessions, like how many errors are coming from the providers, like, you know, how how many requests are dropped, how fast how fast these things are.

I thought you're gonna say daily is too much. Daily is no. And even thinking that it's, like, every three hours, roll up of of key metrics and stats.

Yeah. And, like, one of them was, like, error free sessions and other things like that. That was, like, really important because, you know, especially now with agents which are like multi turn.

I have a tweet about this that was like in 2024, which is that like agents will really only work when we get to not only like the in more intelligent models, but better reliability of the infrastructure providers. Right? Are not inference is not like a database, like update uptime.

Right? So there's still differences between providers. There's still differences between performance and difference uptimes.

And that's why you see things like Open Router being very successful and different gateway products. Because like reliability, you need to switch. They go down all the time.

So long story short, yeah, would we would do like, you know, it was almost like video game style. Like we'd have like all the data coming in all the time. Yeah.

And that allowed I used to joke it's by mood ring, like good day, bad day. So it was very successful for us. I think other teams should adopt that, like, data driven approach.

Yeah. Isn't that weird? It yeah.

I totally agree. It's interesting that that space hasn't been, like, captured as much. Yeah.

Like, I guess maybe now. Actually, I'm interested in this, like, shift to a way into in, like, into knowledge work tasks with coding agents. Okay.

I wonder Using coding agents for non coding tasks? Correct. Do you do you do that personally?

I do. Yeah. Yeah.

What what do do? Well, like I was this summer I was doing like I was helping I was trying to automate some of my dad's workflows and stuff like that. And just like some of his he's got some Excel spreadsheets and for like accounting, like like financial accounting or managerial accounting, I guess.

And yeah, just like point cloud code to that stuff and see what happens. And like, it's it ends up doing Python and and and generates the scripts, and it kinda got off down on the hairs. But it was like, even he saw that it was better at it than the chat client.

That makes sense. Super obvious. They they became kinda obvious.

Yeah. It it felt better. I wonder if he can try Cloud4XL and see if he Yeah.

Yeah. I said. He got me right.

The and then, of course, you got the browser. The I don't know what not the browser. Browser space agents, but not not confused with browsers with agents.

Agent browsers. Agent browsers. Low maintenance.

Poor flexity. Yeah. Everybody's a comment.

That guy. And it's like, is that the better? If that's true, then maybe the the general purpose injection point is there.

Yeah. What do you or I have which have you tried any of the agent browsers? All of them.

All of what do wear? What do wear to?

So you just go on like a you just go on like a like a like a a bender. Yeah. And then you just dump it.

Yeah. It should be as easy to close as markdown, and Chrome isn't isn't that good at the performance side of things yet. And you were working on, some browser comparisons before.

I was. But so I tried to build it in Tory. Okay.

I I legit I'm legitimate hate one of Matt tweet, which is Riz's tweet, like, print it out. You see the tweet, like, there's a tweet from Riz and it's like, no one uses everyone, know, can have that homepage. None all of this stuff is totally useless.

I'll I'll pull it up. Like Yeah. Yeah.

Yeah. I I have I called it out today because we like when we launched. Let me get it right because I gotta do it right.

Hold on. Okay. Was incredible how pretty much the entire GitHub homepage is useless and is 1,300,000 views and 19,000 likes.

And this was 05/02/2025. So hurt. The team the team made improvements, and today they launched a new GetUp homepage, which I'm very proud of.

And they should be really proud of. It's got tasks at this top. It's got recent PRs.

Some stuff is still there, like your recent repositories. I still I think there's more work to do, but it's like really overhauled and they did an amazing job with it. They nailed it.

But you know, more work to do, never done.

this concept of pull requests, which we're all familiar with, You write some commits, you open a PR, and then you merge the PR, and you go about your day. So as you scale like larger organizations, and like you look at your history, and there are people who are like very I'll say like have like near religious beliefs about how to get how did you get right? Rebase versus merch.

There's a crowd that wants to fast forward the repository, so to preserve all the history. And then there's a crowd that wants to squash and and and merge into the I'm few squash. Okay.

Anyway, Facebook and I've I've never worked at Facebook, but in my previous startup before Vercel TurboRepo, I did a lot of research on build systems. And at Facebook, they have not only they have their custom build tools called Buck, they also have a custom file system, and they don't use Git. They use Mercurial, which and then now it's sort of custom, and it's all wired together.

And at at at Facebook, they don't use pull requests. They have a different sort of philosophy. You can it's sort of like the best way to analy like, think about this is like, imagine every PR just had one commit in it.

Yeah. You could branch them and re and the critical thing is you can restack them. And then if you restack or make a change later, like earlier in the stack, than later in the stack, and these stacks are just diffs.

Right? The commits are just diffs, and it's the term stack diffs. You can then collapse them and merge the last one and merge them towards the end.

It just gives you a little bit of a nicer workflow. And it's what people If you work on a mono repo, or you work on a very, very large code base, it's a really, really nice way to work. Especially you've got a system that will re that will automatically restack.

And then if you think even more deeply about it, and get really deeper into the weeds, you can decide which diffs in the stack CI should run against, if fancy. You Okay. Mail.

Like, so early commit messages and There's just always yeah. You could decide like maybe this one doesn't need it, or skip that one, or whatever. And you end up getting these like sort of these groups, these stacks.

And it's really nice from a code review perspective, because when you go to update or you can update different part of the stack, it just makes it a little bit more fluid. And so it's what people want. There are a couple tools out there in the market that do this kind of behavior.

One's called Graphite.

And it's the Azure workflow. And so it's been the top pull request, or sorry, the top feature request. Feature request.

Thank you. At GitHub for years from, from the community's perspective. I don't know at GitHub, but, and then as soon as I joined the first, the first thing I did, I was gonna look this up and Well, not the first thing.

I asked how should I it up better? And it was the top feature request, right? And then I went to go like, okay, investigate, like any good product person would.

And there's been multiple attempts at this internally going back to like 2020. Okay. And there was one very, very, very polished attempt too in 2022.

And it just I don't have all the context, so But there was a pretty good implementation. All of the work was done on the client, and it reintroduced this new concept called stacks outside the pull request into GitHub, and it was a little too risky. It was sort of deemed too risky, too big of a change.

That's just what I was told. So anyway, we had a couple of meetings internally already, and we're trying to weave it into the planning and the roadmap. And so hopefully we'll be able to share more updates soon, but like it's a top of the list known feature.

Again, heard. Yeah. And so, and like we're working on it.

this this really should exist at GitHub. We love GitHub. We use GitHub, but like, come on.

And Well, yeah. Feature requests welcome. Like, my DMs are always open.

It's like Oh, careful. I don't know. A 180,000,000 developers.

I I like 100 whatever. I I am of the philosophy that, like, all feedback is a gift. Like, it's all a signal.

Yeah. And the more signal we can collect, the better decisions we can and truly build this really, really useful website and company like together. And that's gonna be the future.

And if we focus just on that, we're gonna be okay. Yeah. We're gonna be okay.

Alright. Yeah. Well, thanks so much, Jared.

Yeah. It's been a real pleasure. Awesome.

Catching up. Yep. Likewise.

Congrats.