Magnus Hedemark
3a4aad40bf
Spawns a panel of expert agents to debate any question with convergence-aware iteration and typed synthesis output. - PydanticAI + PydanticGraph-based Python package - 5-phase debate protocol: compose → premortem → position → cross-examine (iterative, eval-driven) → synthesis - Convergence detection: confidence dispersion, argument novelty, concession rate — stops when diminishing returns - Typed output schemas (Pydantic models) for every phase - CLI tool with markdown and JSON output modes - Custom persona file support - Bootstrap detection: sys.executable -m pip install fallback - agentskills.io compatible SKILL.md with triggers Signed-off-by: Magnus Hedemark <magnus919@pm.me>
2.7 KiB
Debate Protocol
Phase Structure
Phase 1: Compose
A single LLM call generates N expert personas. The prompt prioritizes diversity of initial position over diversity of expertise. Each persona includes:
- Name
- Career background (one paragraph)
- Specific expertise
- Analytical approach
- Bias or experience they bring to the specific question
At least one agent is structurally skeptical (light red-team). At least one approaches from a fundamentally different cognitive frame.
Phase 2: Premortem
Each agent independently writes how the decision already failed — before any positions are formed. This bypasses positional commitment bias. Agents do NOT see each other's premortems.
The premortem output includes:
- Failure scenario (narrative)
- Root causes
- Early warning signals
Phase 3: Position
Each agent forms an independent position. They see their own premortem (for continuity) but NOT other agents' positions or premortems.
Position output includes:
- Stance
- Reasoning chain
- Confidence score (0-1)
- Key assumptions
Phase 4: Cross-Examination (Iterative)
Each agent reads all other agents' positions and responds. They see:
- Every other agent's stance, reasoning, confidence, and assumptions
- Their own previous round's reflection, concessions, and remaining disagreements
Cross-examination output includes:
- Concessions (where the other agent's reasoning was stronger)
- Remaining disagreements (what's still in dispute)
- Updated position (if changed)
- Updated confidence (if changed)
- Reflection on what they learned
- New evidence needed to close gaps
After each round, convergence detection runs:
- If converged → proceed to synthesis
- If diminishing returns → proceed to synthesis
- If genuine disagreement → proceed to synthesis (with divergence report)
- If more debate needed → run another round (up to max_rounds)
Phase 5: Synthesis
The synthesis combines algorithmic metrics (confidence dispersion, argument novelty) with an LLM-generated narrative. The output preserves the distinction between:
- Pre-positional risks (from premortem — uncontaminated by positional commitment)
- Post-positional concerns (survived cross-examination — tested against alternatives)
Design Principles
- Independent thought first — agents form positions before seeing others'
- Diversity over expertise — different approaches beat more expertise with shared framing
- Pre-mortem before position — surface failure modes before committing to a stance
- Convergence is measured, not assumed — algorithmic stopping conditions prevent premature or interminable debate
- Tension is the output — the synthesis surfaces genuine disagreement, not forced consensus