New Engineering blog: We tasked Opus 4.6 using agent teams to build a C compiler. Then we (mostly) walked away. Two weeks later, it worked on the Linux kernel. Here's what it taught us about the future of autonomous software development. Read more:

I have an idea

Before freaking out read this

But the C compiler has been written before. The LLMs were trained on the C compiler.

> Write extremely high-quality tests > Claude will work autonomously to solve whatever > problem I give it. So it’s important that the task verifier > is nearly perfect, otherwise Claude will solve the > wrong problem. This basically sums up genai coding in a single sentence. It's amazing but ultimately it builds exactly what you tell it to, and what you can competently describe. Much easier to recreate things that are already understood than new things which are not yet understood.

I can "build" a C compiler in one command: git clone

You have agents building C compilers but your support team can't respond to chats or account suspension appeals in over a month?

It also *incorrectly* complies this program without errors! int add(int a, int b) { return a + b; } int main(void) { return add(40); } GCC catches the error - this compiler does not verify argument lengths of functions!

its great stuff for sure , but just under 20k usd , crazy and i dont think any individual claude code user will be ever able to have 2000 parallel claude code sessions even if u lower ur prices , and its most definitely not possible in next 2-3 years imo , hoping i am wrong ofc

Wait… don’t we already have that? So it’s getting really good at rebuilding things humans already built?

Vibe coding a C compiler

"We (mostly) walked away" is doing heavy lifting. The researcher constantly redesigned tests, built CI pipelines when agents broke each other's work. Orchestration > raw intelligence.

Congrats on inventing the first compiler powered by neglect. Next milestone: it files its own bug report, then sues the bug for defamation.

How much token does it cost?

We are watching the shift from writing code to managing context. When the cost of engineering complex systems approaches zero, the only real differentiator left is distribution and the judgment to tell a viable product from more noise.

so; but not cursed

"mostly walked away" doing a lot of heavy lifting in that sentence. curious what the intervention rate actually looked like week 1 vs week 2

If it wasn't obvious yet that software engineering as we know it is over: who's going to pay you $300k/year to write 50-150 lines of code a day when AI can do this?

Sorry this not that impressive or useful, of course it can recreate what it was trained on, I'm sure Opus 4.6 saw all the source code for these compilers during it's training. See if it can create a compiler for a new CPU or a new language, that would be really interesting.

This is terrifying actually. It used to take people months (and their mental sanity) to build something relatively as complex, just 1 year ago. Now we get this with little to no human effort

More large-scale cloning of existing software, cool

never expected my words will come true like this :

Actually Claude helped write this one, and it's better than yours :)

$20,000 & 2 billion token inputs is the real stress testing I believe.

They are literally just using up power at this point. What a pointless demo.

20k to produce a working c compiler that builds the linux kernel feels wild. even a small team of compiler engineers would cost that per day in salary and it would take months not weeks

If you included the Linux Kernel and GNU C Compiler sources in your training set then you are in breach of the license and your model is out of compliance.

After you stole<ahem>learn-from existing FLOSS compilers you mean?

tell your opus 4.6 to finish my medical agentic system that opus 4.5 orchestrated

Since Language models know what a C compiler is they could code it easily. I am just curious can they build something that was never explored or written somewhere in the internet ?

This is wild but I keep coming back to that $20k price tag. You’re right that it’s cheaper than a team, but at what point does the debugging overhead of autonomous agents eclipse just… writing it yourself? The 16-bit x86 workaround feels telling - when the model hits its limits, you still need human intervention to either fix it or architect around it. Still, compiling Linux is no joke.

“It does not have its own assembler and linker; these are the very last bits that Claude started automating and are still somewhat buggy. The demo video was produced with a GCC assembler and linker.” Also, 100k lines of code for C seems like a lot. I guess it’s all the standard libraries as well. None of that seems hard, C is a very simple language. The trickiest part is building the syntax tree.

do you guys still have jobs?

A 100,000-line compiler that builds Linux, written by agents with minimal oversight. The developer role is shifting from code writer to system orchestrator. The question isn't whether agents can code, but whether we can build verification fast enough to keep up.

one researcher and 16 AI agents built a compiler that passes 99% of GCC's torture test suite and runs Doom. GCC took thousands of engineers 37 years. the "(mostly) walked away" does heavy lifting. no human wrote code but the researcher constantly redesigned tests, built CI when agents broke each other's work, and created workarounds when all 16 agents hit the same bug. the work became engineering the environment where AI can write code. you need perfect tests because Claude builds exactly what you specify. As of that... the human role shifted to orchestration and defining what "correct" actually means.

So it didn’t actually build a compiler, it cost 20k and still didn’t actually complete its task It almost sounds like a real programmer