Building a C compiler with a team of parallel Claudes

anthropic ^ | 02 05 2026 | Nicholas Carlini

[yesthatjallen]

I've been experimenting with a new approach to supervising language models that we’re calling "agent teams."

With agent teams, multiple Claude instances work in parallel on a shared codebase without active human intervention. This approach dramatically expands the scope of what's achievable with LLM agents.

To stress test it, I tasked 16 agents with writing a Rust-based C compiler, from scratch, capable of compiling the Linux kernel. Over nearly 2,000 Claude Code sessions and $20,000 in API costs, the agent team produced a 100,000-line compiler that can build Linux 6.9 on x86, ARM, and RISC-V.

The compiler is an interesting artifact on its own, but I focus here on what I learned about designing harnesses for long-running autonomous agent teams: how to write tests that keep agents on track without human oversight, how to structure work so multiple agents can make progress in parallel, and where this approach hits its ceiling.

Enabling long-running Claudes

Existing agent scaffolds like Claude Code require an operator to be online and available to work jointly. If you ask for a solution to a long and complex problem, the model may solve part of it, but eventually it will stop and wait for continued input—a question, a status update, or a request for clarification.

SNIP

---

TOPICS:
KEYWORDS: agents; agentteams; ai; anthropic; arm; autonomous; c; ccompiler; claude; claudes; claudistic; claudistics; linux; llm; nicholascarlini; notjustariverinegypt; parallel; riscv; rust; x86

[yesthatjallen]

Yes. It can run Doom.

[556x45]

Does the compiled image run?

[Dr. Sivana]

If we are going to use AI to make C code, why not use it to make Assembler code? Or would a good compiler be about as efficient?

[grey_whiskers]

What happens when the AI companies have to start charging enough for tokens to pay their actual infrastructure cost?

[posterchild]

I’d be curious to see what it would produce if Claude was tasked with creating its own Context Free Grammar language.

[CodeToad]

“What happens when the AI companies have to start charging enough for tokens to pay their actual infrastructure cost?”

I’ve been wondering that myself. Considering the money being spent on AI data centers, how much work will they actually generate that is billed at a profitable rate?

[dfwgator]

The specs are the code.

Then you can tell the agent to build it in whatever language you want.

So why not assembler?

I’ve never coded in C before, so I told Claude Code to build me a sophisticated application that is a common use case for using C, and document it so I can easily follow what the code is doing, and then told it to create a Guide that walks through the code with thorough instructions on how to build it from start to finish with Visual Studio Code.

[dfwgator]

And I also told it that I am proficient at both Java and Python, so write the doco for someone who knows those languages, and provide reference points.

So basically it’s a technical book that’s personalized for your background.

[Worldtraveler once upon a time]

Welcome to yet more MARKETING. In this case, the YouTube video is for a channel created "Joined Jan 22, 2026."

As to "anthropic," the domain was created 2001-10-02 by Mark Monitor, and the company is in San Francisco. It registered the domain "claude.ai" in 2018-08-04 through the same registrar. It is a company like so many others, incorporated in Delaware and selling stock.

Here are recent "trustees:"

Jason Matheny: CEO of the RAND Corporation Kanika Bahl: CEO & President of Evidence Action Neil Buddy Shah: CEO of the Clinton Health Access Initiative (Chair) Paul Christiano: Founder of the Alignment Research Center Zach Robinson: Interim CEO of Effective Ventures US

How about that Clinton one? With BJ as president and daughter Chelsea as Vice chair of the board? Down the rabbit hole, one finds many trails and many interesting associations. Click. Et cetera.

Anthropic is the maker of Claude, and peddling it in competition with other AI companies.

Skepticism in much is warranted.

[Resolute Conservative]

Another step toward unleashing our future overlord. Nobody ever pays attention and heeds warnings. Human hubris will enslave or kill us all.

[SunkenCiv]

https://en.wikipedia.org/wiki/Anthropic

sidebar(s):

https://www.redshirtjeff.com/shop/p/recompile-linux-shirt
https://www.youtube.com/watch?v=dQ841Pd6YvQ
https://www.youtube.com/@JeffGeerling/videos
https://www.youtube.com/@Level2Jeff/videos

https://www.youtube.com/@ExplainingComputers/search?query=linux
https://www.youtube.com/@ExplainingComputers/videos

[dfwgator]

The wave is coming, you either can try to surf it, or stand still and get swept away. But there’s no stopping the wave.

[cymbeline]

“Does the compiled image run?”

A couple of other questions:

Why do we need another C compiler? C++ is the frontier.

A simple compiler that does no optimization in the generated code would be nothing more than an academic exercise.

[sauropod]

.

[fuzzylogic]

This is quite remarkable.

I’ve 40 years in software development and have been using AI a lot - but I’m clearly not using it to its full potential.

[Lazamataz]

Something I suspect you might be interested in! 😊

[Lazamataz]

A simple compiler that does no optimization in the generated code would be nothing more than an academic exercise.

Absolutely, however, this is a Proof of Concept. It is worthy of my inspection, since I am relatively deep into AI.

[dfwgator]

What about embedded systems?

[catnipman]

“What happens when the AI companies have to start charging enough for tokens to pay their actual infrastructure cost?”

is suspect that’s the plan: first get the populace hooked on free usage and then jack up the charges ... Grok already has several tiers of capabilities, all of which cost money except the lowest tier, which is still currently free ...

[Resolute Conservative]

Luckily, I will expire before the wave come ashore. I have been in software for 20+ years and this latest push with AI is junk. Every time they task me with using it my velocity slows down because I am constantly correcting the AI. So, I turn it off or avoid it, except for looking up simple syntax issues that I have forgotten over the years.

Besides I am not about to teach something how to take my job. I will let another sucker do that. Better yet pull a Jeremiah Johnson and just go live in the woods.