Debate Agents

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](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 → DONEWITHCONCERNS if no convergence). Critic auto-fails synthesis that manufactures consensus, hides outliers, or forces a tiebreaker.

Before Starting

Apply [references/_shared/before-starting-check.md](references/shared/before-starting-check.md) [PLAYBOOK]. Step 0 — Mode resolution: load [references/_shared/mode-resolver.md](references/shared/mode-resolver.md) [PROCEDURE]; standard default; auto-downgrade ≤3-sentence single-decision → fast (debate 2×2 / poll 5×1). --fast does NOT skip structured-output or early-convergence detection. Sub-routine invocations skip steps 1+2 — caller owns problem framing and context. No experience/ read — each invocation is fresh. Session execution profile (single-vs-multi): inherit per references/_shared/execution-policy.md.

Pre-Dispatch

Entry points (standalone vs sub-routine), Warm/Cold Start prompts, sub-routine protocol: [references/procedures/pre-dispatch.md](references/procedures/pre-dispatch.md) [PROCEDURE].

Decision Tree

Mode routing keywords, debate path (A1–A4), poll path (B1–B5), config table, edge cases, costs: [references/decision-tree.md](references/decision-tree.md) [PROCEDURE]. Step prompts in [references/procedures/debate.md](references/procedures/debate.md) and [references/procedures/poll.md](references/procedures/poll.md).

Artifact Contract

• Path (standalone): docs/forsvn/artifacts/meta-debate-agents-[YYYY-MM-DD]-.md (flat v2; dated; immutable per-run)
• Path (sub-routine): none — inline synthesis to caller
• Lifecycle: decision (audit trail; never overwritten)
• Frontmatter: skill, produced_by, version, date, status, stack (=meta), review_surface (html standalone / none sub-routine), decision_state, review_tool, reviewed_at, reviewer, mode, agents, rounds (debate only), provenance
• Sections (debate): Participants / Consensus / Key Disagreements / Recommended Action / Unresolved Risks / Debate Highlights
• Sections (poll): Consensus / Divergences / Outliers / Raw Data / High-Variance Flags
• Consumed by: operator (decision audit trail). No machine consumer in v1.

Full frontmatter + report bodies (debate + poll) + slug convention: [references/report-template.md](references/report-template.md) [PROCEDURE].

Anti-Patterns

[references/anti-patterns.md](references/anti-patterns.md) [ANTI-PATTERN] — re-read before spawning into vague problems, claiming consensus that didn't emerge, forcing a tiebreaker, or writing a sub-routine result to disk.

Durable Rules (protected)

Completion Status

• DONE — debate converged, OR poll yielded clear synthesis
• DONEWITHCONCERNS — unresolved disagreement / split; flagged under Unresolved Risks / Divergences
• BLOCKED — agents couldn't engage (under-specified after Cold Start, scope unclear, rounds exhausted)
• NEEDS_CONTEXT — needs upstream artifact (spec, ICP, architecture); sub-routine returns to caller with what's missing

Next Step

Standalone: operator reviews dated file (Roughdraft / HTML preview), records decision_state. Sub-routine: integrate recommendation into caller's flow.

Chain Position

Standalone or sub-routine. Typical sub-routine callers: prioritize, architect-system, discover.

Worked Example

[references/examples/debate-walkthrough.md](references/examples/debate-walkthrough.md) [EXAMPLE] — 3-agent debate, mind-change, early convergence.

References

• references/playbook.md [PLAYBOOK], references/decision-tree.md [PROCEDURE], references/report-template.md [PROCEDURE], references/anti-patterns.md [ANTI-PATTERN]
• references/procedures/{pre-dispatch, debate, poll}.md [PROCEDURE]
• references/examples/debate-walkthrough.md [EXAMPLE]
• references/_shared/{before-starting-check, mode-resolver, pre-dispatch-protocol}.md

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: hungv47
• Source: hungv47/meta-skills
• License: MIT
• Homepage: https://skills.forsvn.com/

Install and usage instructions live in the source repository linked above.