Agent Room — Stochastic Multi-Agent Discussion

Meta — Multi-perspective panel. Spawns expert agents with distinct stances, surfaces real tradeoffs on high-stakes decisions with no clear winner, recommends a path with reasoning. Standalone or sub-routine.

Core Question: "What do multiple perspectives converge on — and where do they genuinely disagree?"

Why this skill, methodology, principles, when NOT to use, constraint-vs-perspective assignment: references/playbook.md [PLAYBOOK].

Critical Gates

1. Right mode — debate for trade-offs; poll for filtering hallucinations / finding consensus. Default debate.
2. Specific problem — Cold Start before spawning if vague. Don't spawn into ambiguity.
3. Structured output — POSITION/REASONING/PROPOSAL/CONCERNS (debate); ranking/recommendation/binary/scoring (poll). Freeform prose can't be aggregated.
4. Cost — debate 3×3 ≈ $0.30-0.50; poll 10 ≈ $0.30-0.50. Default sonnet; opus ~10× and requires explicit request.

Quality Gate

Synthesis must name real tradeoffs (not collapse to consensus), surface where stances genuinely disagree, recommend an action + reversibility cost, flag unresolved risks honestly (status → DONE_WITH_CONCERNS if no convergence). Critic auto-fails synthesis that manufactures consensus, hides outliers, or forces a tiebreaker.