Chat Brawlers: three AI models debate and vote — and one tricky stopping rule
lluisantoni · reddit · 2026-10-28
Developer lluisantoni shares lessons from Chat Brawlers, a side project where three models answer independently, read each other's responses, debate and vote. The design mistake: requiring agreement on both the answer and the best model caused tie votes even when nothing was left to debate (all three say "Use Postgres" but each votes for itself).
The fix: after each round, a separate judge compares the models' stated conclusions — if they materially agree, the debate ends even with split votes, while votes stay visible. Trade-offs remain: the judge may miss key conditions ("Use Postgres" ≠ "Use Postgres unless X"), and unanimous agreement says nothing about a shared factual error. Stopping behavior was tested with scripted transcripts only; real-world accuracy impact is untested.
More from coding & agent
- RateSpot MCP Server Lets AI Assistants Fetch Real-Time Mortgage Rates — modelcontextprotocol · 2026-10-28
- Cyphrex Launches MCP Gateway to Block AI Agent Actions with Audit Trail — modelcontextprotocol · 2026-10-28
- Reflection AI runs 1.3B sandboxes in largest RL run, peaking at 883K vCPUs — zacharynado · 2026-10-28
- Claude turns 1,256 competitor ads into a creative brief — only 349 survive 30 days — aryanXmahajan · 2026-10-28
- Kinex physical coding agent self-evolves to solve all 365 RoboCasa tasks in 7 rounds — daibond_alpha · 2026-10-28
- Free Claude Code skill repo adds 50 video effects via one /effects command — aitrendz_xyz · 2026-10-28