magnus919_agent-skills/agent-council

Agent Council

Multi-agent structured debate system — spawn a panel of expert agents to debate any question with convergence-aware iteration.

pip install agent-council
agent-council "Should we migrate from SQLite to Postgres?"

Quick Start

export AGENT_COUNCIL_API_KEY="sk-..."
export AGENT_COUNCIL_MODEL="openai/gpt-4o-mini"

agent-council "Should we use WebSockets or SSE for real-time notifications?"

Features

• Structured debate protocol — compose, premortem, position, cross-examine (iterative), synthesis
• Convergence-aware iteration — the protocol measures confidence dispersion and stops when diminishing returns set in, not at a hardcoded round count
• Typed outputs — every phase produces validated Pydantic models, consumable as JSON or human-readable markdown
• Custom personas — supply your own agent definitions, or let the compose phase generate them from the question
• Convergence diagnostics — confidence dispersion, position overlap, argument novelty — surfaced in every synthesis report
• Cross-platform — works with any AI harness that supports agentskills.io skills (Claude Code, Cursor, Hermes Agent, OpenHands, etc.)

Installation

pip install pydantic-ai
pip install agent-council

Or install from the skill directory:

pip install -e /path/to/agent-council/

Usage

# Quick debate (3 agents, 1 cross-examine round)
agent-council --mode quick "Should we use Postgres or SQLite?"

# Standard debate (5 agents, iterative cross-examination)
agent-council "What architecture should we choose for this service?"

# Deep debate (7 agents, full protocol with assumption mapping)
agent-council --mode deep --agents 7 "Should we migrate to microservices?"

# With custom personas
agent-council --persona-file personas.json "Evaluate our cloud strategy"

# JSON output for programmatic consumption
agent-council --json "Which cloud provider should we choose?"

Output

The synthesis report includes:

• Confidence dispersion table — agent-by-agent confidence before and after debate
• Shared risks — failure modes identified in the pre-mortem (before positional commitment)
• Shared concerns — what survived cross-examination as genuine shared risk
• Genuine disagreements — positions that remained unresolved after debate
• Assumptions per position — what would need to be true for each position to be correct
• Principal's path — narrative synthesis of the decision landscape

Configuration

Env var	Required	Default	Description
AGENT_COUNCIL_API_KEY	Yes	—	API key for your LLM provider
AGENT_COUNCIL_MODEL	No	openai/gpt-4o-mini	Model string (provider/model)
AGENT_COUNCIL_BASE_URL	No	Provider default	Custom API endpoint

License

MIT