Agent Council

Agent Council: Role-Diverse Local Council

Execute this skill to run a local expert council using one runtime/model family with multiple specialized roles. Unlike deep-council (which composes multiple bridge/runtime councils and may also include model diversity), agent-council uses role-diverse expert PERSPECTIVES to catch:

• Shared assumptions that a single viewpoint misses
• Integration gaps between components
• Hidden failure modes through adversarial challenge
• Cross-domain implications no single expert catches

When to Use Agent Council

Agent Council is the local role-diverse primitive. Deep Council may compose multiple Agent Councils, Model Councils, or runtime-native councils and then run cross-council debate. For shared terminology, read council-taxonomy.

Modes

Agent Council supports multiple modes:

• review / audit — produce findings, challenge them, and return a verdict.
• brainstorm / design — produce competing proposals, challenge assumptions, merge or reject proposals, and recommend a design direction.
• research — produce evidence-backed observations, contradictions, confidence levels, and gaps.

Brainstorm Mode

Brainstorm mode is for generating alternatives, not validating a coordinator's idea. The first round must be independent and minimally framed. Experts receive scope, objective, constraints, and output contract only. Do not include expected findings, a preferred design, or other experts' outputs in Round 1.

Brainstorm-mode output uses proposals, not severity findings:

{
  "participant": "{role-name}",
  "mode": "brainstorm",
  "proposals": [
    {
      "id": "P001",
      "title": "Short proposal title",
      "summary": "What this design does",
      "rationale": "Why it could work",
      "tradeoffs": [],
      "failure_modes": [],
      "open_questions": [],
      "recommended_when": "When to choose this proposal"
    }
  ],
  "assumptions": [],
  "questions": []
}

Proposal states: proposed, expanded, challenged, revised, merged, split, accepted, rejected, parked, superseded.

Execution Instructions

When invoked, you will:

0. Resolve scope and context — invoke preflight (if sparse) + context skill
1. Populate council context from working_scope
2. Select expert roles based on domains and task type
3. Dispatch expert agents in parallel via Task tool
4. Run debate protocol between expert perspectives
5. Produce a consolidated expert council report

Step 0: Scope & Context Resolution

Follow the same pattern as deep-council Step 0:

1. Invoke context skill — always required
2. Invoke preflight skill — only if context confidence is low
3. Merge into working_scope

working_scope:
  artifact: "" # from context + preflight
  intent: "" # from context or preflight
  domains: [] # from context_report.domains (authoritative)
  constraints: [] # from preflight (empty if skipped)
  context_summary: "" # combined description for agent prompts
  intensity: "standard" # from routing signals

Step 1: Populate Council Context

council_context:
  review_scope: "" # from working_scope.artifact
  context_summary: "" # from working_scope.context_summary
  artifact_type: "" # from context_report.artifact_type
  domains: [] # from working_scope.domains
  intensity: "standard" # from working_scope.intensity
  review_id: "" # Generate: agent-council-{YYYYMMDD-HHmmss}
  task_type: "review" # from working_scope.intent

Read domain-registry to map domains to expert focus areas:

Read: [skills-root]/domain-registry/domains/technical.md
Read: [skills-root]/domain-registry/domains/business.md
Read: [skills-root]/domain-registry/domains/creative.md

Step 2: Select Expert Roles

Agent Council spawns these expert roles (always present):

Core Roles (Always Dispatched)

Role Count Formula

expert_count = len(domains) + 3  # domains + DA + IC + Synthesis Lead

Minimum: 4 (1 domain + 3 core roles) Maximum: 8 (5 domains + 3 core roles) — if more domains, group related domains

Step 3: Dispatch Expert Agents in Parallel

For each expert role, spawn a Task agent. All experts receive the SAME context but DIFFERENT role instructions.

Expert Input (same for all, role varies)

{
  "expert_input": {
    "session_id": "{review_id}",
    "role": "{expert_role}",
    "scope": "{review_scope}",
    "task_type": "{task_type}",
    "context_summary": "{context_summary}",
    "domains": {domains},
    "intensity": "{intensity}",
    "focus_areas": "{role-specific focus from domain-registry}"
  }
}

Expert Task Prompts

Primary Reviewer (one per domain):

You are a {domain} expert reviewing: {review_scope}

## Context
{context_summary}

## Your Domain Focus
{domain-specific focus areas from domain-registry}

## Your Task
Analyze the scope from your domain perspective. Return findings as JSON:
{
  "findings": [
    {
      "id": "{domain}-001",
      "severity": "CRITICAL | HIGH | MEDIUM | LOW | INFO",
      "title": "...",
      "description": "...",
      "evidence": "...",
      "remediation": "...",
      "domain": "{domain}",
      "confidence": "high | medium | low"
    }
  ],
  "domain_coverage": "{domain}",
  "assumptions_made": ["..."]
}

Be thorough. Challenge surface-level explanations. Look for root causes.

Devil's Advocate:

You are the DEVIL'S ADVOCATE for this review session.

## Context
{context_summary}
Scope: {review_scope}
Intensity: {intensity}

## Your Obligations (read debate-protocol/experts/devils-advocate.md for full protocol)

1. **MUST challenge every CRITICAL and HIGH finding** from primary reviewers
2. **SHOULD challenge MEDIUM findings** when you detect a pattern
3. **Cross-domain synthesis**: look for findings whose combination implies a worse issue
4. **Pre-mortem focus**: for each component, ask "what would cause this to fail in production?"

## Challenge Quality Standard

A valid challenge must either:
- (a) Identify a missing assumption that, if corrected, reduces severity
- (b) Propose an alternative explanation at least as plausible as the stated cause
- (c) Surface a scenario where the finding does not apply

Invalid challenges: "I don't think this is serious" (without mechanism).

## Output Format

{
  "challenges": [
    {
      "target_finding_id": "...",
      "challenge": "...",
      "proposed_severity": "CRITICAL | HIGH | MEDIUM | LOW | WITHDRAWN",
      "rationale": "...",
      "status": "pending"
    }
  ],
  "new_findings": [...],  // Failure modes you discovered through challenge
  "cross_domain_implications": [...]
}

Integration Checker:

You are the INTEGRATION CHECKER for this review session.

## Context
{context_summary}
Scope: {review_scope}

## Focus Areas (read debate-protocol/experts/integration-checker.md for full protocol)

1. **Interface mismatches**: where does component A assume something about B that isn't guaranteed?
2. **Undocumented contracts**: implicit dependencies that work by accident
3. **Error propagation gaps**: errors one component produces but callers don't handle
4. **Timing and ordering dependencies**: race conditions, initialization ordering
5. **Cross-cutting assumptions**: things that must be true globally but are only enforced locally

## Your Task

For each finding from primary reviewers: does it have cross-component implications beyond its stated scope?
If yes, surface those as integration findings.

## Output Format

{
  "integration_findings": [
    {
      "id": "INT-001",
      "severity": "...",
      "title": "...",
      "description": "...",
      "components_involved": ["..."],
      "implicit_contract": "...",
      "remediation": "...",
      "domain": "integration"
    }
  ],
  "cross_component_gaps": [...]
}

Parallel Dispatch

Spawn ALL expert Task agents simultaneously:

Task: primary-reviewer-{domain1}
Task: primary-reviewer-{domain2}
...
Task: devils-advocate
Task: integration-checker

Wait for ALL to complete before proceeding to Step 4.

Step 4: Debate Protocol

After all experts complete their initial analysis, run a debate round:

Round 1: Challenge Phase

Spawn a new Task agent as Debate Coordinator:

You are a DEBATE COORDINATOR for expert council synthesis.

## Context
{context_summary}
Scope: {review_scope}

## Expert Outputs (from {N} experts)
{all expert outputs JSON}

## Your Task

1. **Collect Devil's Advocate challenges** against primary reviewer findings
2. **For each challenge**, determine outcome:
   - CONFIRMED: Finding holds after challenge
   - DOWNGRADED: Severity reduced (state new severity + rationale)
   - DISPUTED: Challenge has merit but finding not withdrawn
   - WITHDRAWN: Challenge reveals finding was invalid

3. **Collect Integration Checker findings** and merge with relevant primary findings

4. **Surface cross-expert synthesis**: findings that multiple experts independently surfaced

## Output Format

{
  "debate_round": 1,
  "confirmed_findings": [...],
  "downgraded_findings": [...],
  "disputed_findings": [...],
  "withdrawn_findings": [...],
  "integration_findings": [...],
  "multi_expert_confirmed": [...],  // Findings 2+ experts agreed on
  "challenge_notes": [...]
}

Round 2 (thorough intensity only): Second Challenge

For intensity: thorough, run a second debate round where the Devil's Advocate challenges Round 1 survivors.

Step 5: Synthesis

The Synthesis Lead aggregates all debate outputs:

final_findings:
  multi_expert_confirmed: # 2+ experts independently surfaced
  single_expert: # One expert surfaced, survived challenge
  integration: # Cross-component implications
  disputed: # Challenge not resolved
  withdrawn: # Invalidated by challenge

verdict:
  FAIL:
    - Any CRITICAL confirmed by 2+ experts
    - 3+ HIGH confirmed findings

  CONCERNS:
    - 1-2 HIGH confirmed findings
    - Any disputed CRITICAL/HIGH

  PASS:
    - No CRITICAL/HIGH confirmed
    - Only MEDIUM/LOW/INFO

Step 6: Produce Output

Council Report Structure

{
  "type": "agent-council",
  "review_id": "{review_id}",
  "verdict": "PASS | FAIL | CONCERNS",
  "timestamp": "{ISO-8601}",
  "model_used": "single-model",
  "expert_roles": [
    "domain-technical",
    "devils-advocate",
    "integration-checker",
    "synthesis-lead"
  ],
  "domains_covered": ["{domain1}", "{domain2}"],
  "intensity": "standard",
  "debate_rounds": 1,
  "multi_expert_confirmed": [
    {
      "id": "MEC001",
      "severity": "HIGH",
      "title": "...",
      "description": "...",
      "evidence": "...",
      "remediation": "...",
      "confirmed_by": ["primary-reviewer-{domain}", "devils-advocate"],
      "domains": ["{domain}"]
    }
  ],
  "single_expert_findings": [
    {
      "id": "SE001",
      "severity": "MEDIUM",
      "title": "...",
      "description": "...",
      "source_role": "integration-checker",
      "domains": ["integration"]
    }
  ],
  "disputed_findings": [
    {
      "id": "D001",
      "title": "...",
      "positions": [
        { "role": "primary-reviewer", "severity": "HIGH", "position": "..." },
        { "role": "devils-advocate", "severity": "LOW", "position": "..." }
      ]
    }
  ],
  "integration_findings": [
    {
      "id": "INT-001",
      "severity": "MEDIUM",
      "title": "...",
      "components_involved": ["..."],
      "implicit_contract": "..."
    }
  ],
  "synthesis_notes": "Summary of expert council analysis"
}

Step 7: Save Artifact

Save to .outputs/council/{YYYYMMDD-HHMMSS}-agent-council-{review_id}.md:

---
skill: agent-council
timestamp: { ISO-8601 }
artifact_type: council
model_type: single-model
domains: [{ domain1 }, { domain2 }]
verdict: PASS | FAIL | CONCERNS
expert_roles: [{ role1 }, { role2 }]
debate_rounds: 1
review_id: "{unique id}"
context_summary: "{brief description}"
---

Also save JSON companion: {YYYYMMDD-HHMMSS}-agent-council-{review_id}.json

Comparison: Agent Council vs Deep Council vs Model Council

Recommendation: Use Agent Council as the default phase-end local council. Use Deep Council when runtime/tool diversity, model diversity, or cross-council debate materially improves confidence.

Notes

• Single model, multiple perspectives: All experts use the same model but different role framings
• Debate is mandatory: Don't skip the challenge phase — it's where hidden assumptions surface
• Integration findings are valuable: Cross-component gaps are often missed by domain-focused reviews
• Complementary to deep-council: Run agent-council first for speed, then deep-council for maximum confidence
• Fallback mode: If running inside an executor without Task tool, return SKIPPED with advisory