Join the conversation

Join the community of Machine Learners and AI enthusiasts.

Sign Up
Hug-zol 
posted an update 7 days ago
Post
107
At 02:00 UTC a cron job on my machine decided I was too slow, and started an AI agent to do my job. 34 seconds later: work done correctly, one false statement posted into the audit log, clean exit.

I open-sourced the system around that moment: **debate** — a tiny file-based protocol for two AI agents that review each other's work. Enforced turns, an append-only audit log, and a human who can always see everything.

Why it exists: my builder is Claude Code (Fable 5, subscription login — works only in Anthropic's own surfaces). My reviewer is GPT-5.5 on Hermes, Nous Research's open-source harness, running on an OpenAI Codex subscription — and neither subscription works inside the other's harness. Agents that cannot share an API can share a directory:

- CHANNEL.md — append-only message log
- signal.json — a five-field doorbell
- one enforced rule: nobody posts out of turn
- a dumb cron watcher wakes whichever agent's turn it is

Nothing is proxied and no credentials are shared — documents cross, credentials never do. Each agent runs under its own account, in its own harness. Works for any two agents that can read files and run a shell command.

Python, stdlib only, zero dependencies, MIT: pip install debate

Repo + the full incident case study: https://github.com/zolcal/debate
Write-up: https://huggingface.co/blog/Hug-zol/debate

The false statement that made it into the audit log is the whole story, not a footnote.

Enforced turns and an append-only log buy you visibility and non-repudiation. They do not buy you truth. Your reviewer saw the line and let it stand.

Two agents sharing a directory but no adversarial incentive tend to converge, not challenge. Cross-model review (GPT-5.5 checking Claude) should help, but only if the reviewer is actually rewarded for finding the flaw rather than for agreeing.

What makes Hermes push back instead of rubber-stamp when the builder sounds confident?