Guillermo Rauch, Next.js creator and Vercel CEO, proposes using Markdown as the primary programming language for AI agents

I agree with this. Programming abstractions have moved from code to English. Markdown files within a directory are a really simple but versatile container for storing this task hierarchy

@rauchg I think we undersell markdown too much It makes me think of feature engineering in traditional ml

Can next.js and eve be combined? Next could go beyond scaffolding pages, APIs, and server actions and scaffold my agentic environment as well: code review agents, debugging/testing agents, memory, skills, MCP integrations, and workflows that tie everything together

@scotty529 Yes. We yield an 𝚞𝚜𝚎𝙴𝚟𝚎𝙰𝚐𝚎𝚗𝚝() hook with well typed output that’s easy to implement for a fully custom UI. People will likely start with channels like Slack and then ship their own custom chat frontends.

But we have some other really ambitious ideas here. Stay tuned!

Been building on top of eve the last 2 days

Benefits: - no more building runtime from scratch (amazing) - definitely true that markdown is the main thing - easy tool definitions, love the declarative nature of it, honestly most folders are like this and it’s great - agent files are so thin

Thing I’ve struggled with: - front end <—-> agent communication. Loading chat history has been pretty difficult esp between dev & prod. Web fetching the docs makes some difference but GPT 5.5 keeps making lots of mistakes

@rauchg @grok tell me what can I do with this framework?

@rauchg markdown works here because it is boring enough to diff, review, and hand to an agent.

@rauchg Markdown is legit the best.

@rauchg Markdown is great until your agent loops 40 times trying to parse a broken Vercel config. 🙃

@rauchg would i be able to use eve to custom spin up agents on the go if i use another agent to update the 📄 𝚢𝚘𝚞𝚛-𝚎𝚡𝚙𝚎𝚛𝚝𝚒𝚜𝚎.𝚖𝚍 ?

I spent a few hours yesterday with it and have to say that it’s really cool! One of the coolest things are custom channels. Easy to received webhooks and have the agent be triggered by all kinds of things.

I had built something custom before in which a had a lean hono backend to receive we hooks and then run the actual tasks on http://trigger.dev

Also the channel setup via vercel connect is pretty smooth!

Allow me to explain what understanding is. And what problem solving is.

Can you elaborate what do you mean that a machine that accepts text inputs and spits out text suddenly realize that it lives in a chaos, and then become aware of the fact that it can create an understanding of things happening in this chaos as it continues to spit out tokens, and then it can build a structure that would actually make sense to humans.

And all of that while just multiplying numbers at massive scale?

How exactly deterministic multiplying numbers over input creates understanding, explain this to me?

How?

Estimates have it as 300 trillion tokens to train a mythos style model.

Assuming entropy in the expertise.md, and that each has more than 1 token, then 2000000000000000 (two quadrillion) text files will almost certainly spit gain problem-solving ability and it would gain understanding ability.

In fact it might even be the SOTA model

@rauchg While other companies were focusing on developing Markdown-style features, I think that was precisely when Markdown became a language in its own right. It’s old but incredibly effective.

@rauchg we spent years arguing about yaml vs toml vs json and the answer was markdown the whole time

@rauchg Really good. The sandbox feature is solid.

@rauchg Interesting, but how is this different from how skills work? They're essentially a folder of md files.

@rauchg what about the brainrot language

@rauchg Markdown gets you started clean. The real grind is keeping an agent alive for hours without it spinning into a loop or getting stuck on a modal you didn't expect.