Open-Source Tool Makes LLMs Debate Until Consensus
trekhleb · reddit · 2026-07-17
Based on Karpathy's llm-council concept, a developer has released Yesbrainer, an open-source tool that runs entirely in the browser. It supports Bring Your Own API Key (BYOK) and connects directly to major LLMs.
Core Mechanism:
Unlike the original's single-round ranking, this tool introduces a consensus mode where models reach a conclusion through multi-round debates:
- All participating models provide independent answers first.
- An arbiter model evaluates if the answers are converging. If not, it extracts the points of divergence for the next round.
- Models re-answer after seeing their peers' anonymized arguments. Tags are shuffled each round to prevent brand identification.
- The loop continues until a consensus is reached or the round limit is hit, ultimately outputting the agreed-upon view and any remaining conflicts.
Real-World Test:
In a test asking to "choose a third language for an 8-year-old bilingual child," Claude initially insisted on French, while GPT and Gemini chose Chinese. During the second round of debate, GPT and Gemini were persuaded by arguments regarding fluency expectations and the importance of immersive environments, both switching their choice to French.
More from coding & agent
- An MCP server signs every AI agent tool call into a verifiable Merkle chain — Funky_Chicken_22 · 2026-07-22
- Claude Code skill uses 10 Markdown rules to make outputs ADHD-friendly — alex_verem · 2026-07-22
- A Firecracker-based platform says it can host 6,000 AI agents on one 256 GB server — maritime_sh · 2026-07-22
- A better path to agent autonomy is running waves, finding friction, and iterating — JnBrymn · 2026-07-22
- AI agent designers map the visual and tonal cues behind companionship products — Unlikely-Platform-47 · 2026-07-22
- Coding agents are heading toward an AI-writes, AI-reviews, human-approves workflow — aftahi_ai · 2026-07-22