Even with the strongest guardrails, no llm can be left unsupervised. And believing an llm can supervise itself is madness. For every agent you deploy as a “team” you will have multiplied diminishing returns.

I’ve iterated the process hundreds of times and even with a strong specification and tests, an llm can meet the spec and pass all the tests and still build the wrong thing.

This multi-agent workflow idea is worse than web3.

But it’s a YC thing so someone will make money.

I also don't believe in this multi-agent swarm thing. I just use Claude Code "manually" and this is still useful for that.