Recoleta
GitHub
Source note

Conclave – make LLMs debate each other before they respond

Authors
andngd
Source
Hacker News
Published
Collection
Software Intelligence
Multi Agent Software EngineeringLLM DebateCode AgentsAutomated Software ProductionDeveloper Tools
Open originalMarkdown source

Summary

Conclave adds a structured multi-LLM debate engine to OpenCode so several models can discuss a software task before one response or implementation is chosen. It targets code work where a single model may miss architecture flaws, edge cases, or security issues.

Problem

  • Single-model coding agents can miss design flaws, edge cases, and security risks before code is written.
  • Teams that already pay for Claude Code, Gemini CLI, Codex, or local models may lack direct API access for multi-model workflows.
  • Multi-model debate adds latency and cost, so the tool is aimed at complex tasks rather than simple prompts.

Approach

  • Conclave modifies OpenCode with a team debate engine; the page says about 12 files were changed while OpenCode keeps providers, agents, git support, tools, and the TUI.
  • Models debate in structured rounds using LEAD, SUPPORT, ALIGN, BUILD, and CHALLENGE signals.
  • A winner is selected by endorsement score after the debate.
  • Users can mix CLI-authenticated and API models, including Claude Code, Gemini CLI, Codex, OpenAI, Anthropic, DeepSeek, Google, NVIDIA, Groq, and Ollama.
  • Each model gets a debate thread sized to its context limit; large-context models can see more context, while smaller models get signal summaries.

Results

  • The excerpt reports no benchmark, dataset, or human-evaluation results showing output-quality gains over a single LLM.
  • Implementation scope: about 12 OpenCode files modified to add the debate engine.
  • Cost and latency scale with team size and rounds: 3 models debating for 3 rounds make 9 API calls per user message.
  • Cost example: each team member makes an independent call, so a 3-model team costs about 3x a single-model call unless free CLI or local models are used.
  • Context claim: the tool can pair a 1M-token DeepSeek model with a 128K-token Gemini Flash model by truncating or summarizing the debate thread per model.
  • Maturity claim: the project is weeks old; debate, team persistence, and provider connections work, while context optimization, live streaming, and Breaking Teams are roadmap items.

Link