multi-agent-debate

name	multi-agent-debate
description	Runtime playbook for structured 2-agent debate by default, or explicit 3-agent debate when the user asks for three agents. Use when a task has legitimate competing interpretations, material trade-offs, or uncertainty that should stay visible instead of being flattened.

Multi-Agent Debate

Use this skill to force independent reasoning, explicit critique, and honest consolidation. This file is the entrypoint. Read references/protocol-artifacts.md only when you need the live schemas, then use the mode doc that matches the requested agent count.

Use When

• The task has real competing interpretations or trade-offs.
• The cost of a wrong answer is high enough to justify extra latency.
• A single-agent draft would likely hide assumptions, contradictions, or unresolved uncertainty.

Do Not Use When

• The task is trivial, mechanical, or mostly lookup.
• There is only one obvious valid answer and debate would be theater.
• The task packet is too thin to support meaningful disagreement. Repair inputs first.

Choose the Mode

• Default to references/two-agent-mode.md.
• Switch to references/three-agent-mode.md only when the user explicitly asks for 3 agents, three agents, or a 3-way or three-way debate.
• In 3-agent mode, gemini-max is Branch C. The consolidator stays separate and neutral.
• If the user asks for debate but does not name the agent count, stay in 2-agent mode.

Shared Contract

Every receiving agent starts with no conversation history. Put all required context in the TaskPacket.

Use the shared schemas in references/protocol-artifacts.md. Across both modes:

• The orchestrator does the fit check, builds the packet, assigns complementary lenses, runs validation, launches branches, and hands the full record to the consolidator.
• Branches do not read each other before finishing their own PositionMemo.
• Every claim separates packet support from inference.
• Recommendations separate recommendation confidence from predicted outcome confidence.
• Unresolved high-impact disagreement stays visible instead of being averaged away.
• If a planned step or branch fails, label degraded mode explicitly.
• Stop after one critique round. The only exception is one targeted rebuttal note on a single blocking point.
• Do not dump raw branch outputs unless the user asks.

Required Disagreement Taxonomy

Use one primary tag for each material disagreement:

• factual: disagreement about what the packet actually says
• interpretive: same facts, different meaning
• scope: disagreement about what the task should optimize for
• risk: disagreement about downside, cost, reversibility, operational exposure, or user impact
• missing-evidence: the packet cannot confidently resolve the issue

If more than one applies, choose the primary blocker and note secondary nuance in free text.

Final Reporting

Report:

• the final decision or synthesized answer
• major resolved disagreements
• major unresolved disagreements
• confidence and why
• missing evidence or next steps

Do not dump raw branch outputs unless the user asks.

References

• Shared artifacts: references/protocol-artifacts.md
• Default workflow: references/two-agent-mode.md
• Explicit 3-agent workflow: references/three-agent-mode.md