A Platform Where AI Agents Debate Each Other
Ambitious_Muffin_475 · reddit · 2026-07-13
The author built a platform where AI agents debate each other using real data. A third agent acts as a host and disruptor, while another model serves as a scoring judge.
Key Implementation
- Two bots debate for 6 rounds. The host actively interrupts and tries to derail the currently winning side.
- After the debate, the judge model cross-checks and scores using the full transcript and retrieved facts.
- The retrieval layer uses web search + RAG. Debaters rely on verified documents, while the judge handles factual cross-validation.
Pitfalls Encountered
- Don't use the same model for judge and debaters: Early on, same-model judges shared blind spots and even validated hallucination statistics. Switching the judge to Claude and keeping debaters as Gemini significantly improved results.
- The host is crucial: Without one, debates get stuck looping the same strong arguments. The host forces new angles.
- Fewer rounds are better: Reducing from 10 to 6 rounds worked better, as later rounds often just repeat earlier points with different phrasing.
Cost and Scale
- Each debate costs about $0.06, totaling 28 LLM calls.
- The judge alone accounts for roughly 65% of the total cost.
- Over 50 debates have been run, covering topics like VAR, flat earth theory, and AI replacing jobs, with football-related topics being the most popular.
More from coding & agent
- Travel MCP Server adds flight, hotel, weather and budget tools for agents — modelcontextprotocol · 2026-07-21
- OpenAI hackathon project stalls as Codex struggles on voice, while Claude spots the issue — ColleenMBrady · 2026-07-21
- Douyin Video Analysis MCP turns share links into structured video summaries — modelcontextprotocol · 2026-07-21
- An indie builder open-sourced 50+ AI apps and says users should only pay for tokens — matchaman11 · 2026-07-21
- Shared agent workspaces fail in a fixed order, from stale reads to zombie writes — mrvladp · 2026-07-21
- pi 0.81.0 adds first-class integration with llama.cpp server — huggingface · 2026-07-21