Skip to content

agent-sh/debate

debate

Structured multi-round debate between two AI tools - one proposes, one challenges, and an orchestrator delivers a verdict.

Why

A single AI tool gives you one perspective. When the stakes are high - architecture decisions, schema design, migration strategies - you want adversarial pressure-testing. debate pits two AI tools against each other in a structured format with enforced rules: the challenger must find flaws before agreeing, both sides must cite evidence, and the orchestrator must pick a winner.

Use cases:

  • Stress-test an architecture decision before committing (microservices vs monolith, event sourcing vs CRUD)
  • Get adversarial review of a proposed schema or API design
  • Compare two approaches with structured arguments rather than vibes
  • Force a devil's advocate perspective on a plan you are leaning toward

Installation

agentsys install debate

Requires at least two AI CLI tools installed. See consult for individual tool install commands.

Quick Start

# Codex argues for, Gemini challenges - 2 rounds about microservices
/debate codex vs gemini about microservices vs monolith

# Thorough 3-round debate with explicit flags
/debate "Should we use event sourcing?" --tools=claude,gemini --rounds=3 --effort=high

# Include the current git diff as context for both tools
/debate codex vs gemini about this refactoring approach --context=diff

How It Works

  1. Parse - extracts topic, tools, rounds, and effort from natural language or flags
  2. Detect - scans PATH for installed AI CLI tools; requires at least two
  3. Resolve - if parameters are missing, presents an interactive picker (proposer, challenger, effort, rounds, context)
  4. Debate - executes rounds sequentially. Each round:
    • The proposer argues for the topic (round 1) or defends against challenges (round 2+)
    • The challenger identifies flaws, missing considerations, and alternatives
    • Both must support claims with evidence; unsupported claims are flagged
  5. Verdict - the orchestrator (Opus) reads all exchanges and delivers a structured synthesis: winner, agreements, disagreements, unresolved questions, and an actionable recommendation

The challenger operates under adversarial rules - it must lead with what is wrong or missing before acknowledging agreements, and cannot agree with unsupported claims.

Usage

Natural language

/debate codex vs gemini about microservices vs monolith
/debate with claude and codex about our auth implementation
/debate thoroughly gemini vs codex about database schema design
/debate codex vs gemini 3 rounds about event sourcing

Explicit flags

/debate "Redis vs PostgreSQL for caching" --tools=codex,opencode
/debate "Should we use event sourcing?" --tools=claude,gemini --rounds=3 --effort=high
/debate "Performance optimization" --tools=codex,gemini --context=diff

Flags

Flag Values Default Description
--tools TOOL1,TOOL2 interactive Proposer and challenger (comma-separated)
--rounds 1-5 2 Number of back-and-forth exchanges
--effort low, medium, high, max interactive Reasoning depth for both tools
--context diff, file=PATH, none none Codebase context shared with both tools
--model-proposer any model identifier from effort Override proposer model
--model-challenger any model identifier from effort Override challenger model

Debate output format

Each debate produces a structured result:

  • Rounds - proposer opening, challenger response, and subsequent exchanges displayed as they complete
  • Verdict - winner with cited evidence from the debate
  • Debate quality - rated on genuine disagreement, evidence quality, and challenge depth
  • Agreements - points both sides converged on, with supporting evidence
  • Disagreements - points where they diverge, with each side's position
  • Unresolved questions - gaps neither side addressed
  • Recommendation - actionable next step (the orchestrator must pick a direction)

Requirements

  • agentsys runtime
  • Node.js (for ACP transport detection)
  • At least two supported AI CLI tools on PATH (claude, gemini, codex, opencode, copilot, or kiro)

Related Plugins

  • consult - single-tool consultation for quick second opinions
  • agentsys - plugin runtime and orchestration

License

MIT

About

Structured multi-round debate between AI tools with proposer/challenger roles and verdict

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Packages

 
 
 

Contributors