panel-of-experts

name	panel-of-experts
description	Runs multiple expert agents in parallel on the same problem, normalizes their outputs, detects consensus or disagreement, and synthesizes a final recommendation. Trigger — "get multiple opinions", "panel of experts", "expert consensus", "compare approaches", "run experts in parallel", "multi-perspective review", "which approach is best", "synthesize expert outputs". Skip — trivial decisions with one obvious answer, tasks where a single specialist suffices, real-time latency-critical paths that cannot afford parallel fan-out.
license	Apache-2.0
compatibility	{"clients":["openai-codex","gemini-cli","opencode","github-copilot"]}
metadata	{"owner":"codex","domain":"panel-of-experts","maturity":"draft","risk":"low","tags":["panel","experts","consensus","synthesis","parallel"]}

Purpose

When a problem benefits from multiple perspectives — architecture decisions, code review, risk assessment, debugging — launch N expert agents in parallel with differentiated prompts, collect their independent outputs, detect where they agree and disagree, and synthesize a single actionable recommendation with confidence levels.

When to use

• A decision has high stakes and benefits from independent expert review (architecture, security, design).
• The team wants to compare multiple implementation approaches before committing to one.
• A code review or audit should cover different concerns (correctness, performance, security) in parallel.
• The problem is ambiguous and you need to surface disagreement rather than follow one opinion.
• A previous single-agent attempt produced a low-confidence or controversial result.

Do NOT use when

• The task has a single correct answer that one specialist can produce (e.g., fix a syntax error).
• Latency budget is too tight for parallel fan-out and synthesis (under 10 seconds total).
• Fewer than 2 meaningfully different expert perspectives exist for the problem.
• The task is execution, not analysis — building code rather than evaluating options.

Operating procedure

1. Define the problem statement. Write a clear, self-contained problem description that each expert will receive. Include all relevant context, constraints, and evaluation criteria.
2. Select the expert panel. Choose 2-5 experts based on the problem domain. For each expert, write a one-line role description. Output as: | Expert | Specialty | Perspective Focus |. Examples: security-auditor, performance-engineer, maintainability-reviewer, domain-specialist.
3. Differentiate the prompts. For each expert, craft a prompt that shares the same problem statement but adds role-specific instructions. Each prompt must explicitly say: "Focus on [X]. Do not attempt to cover all aspects — other experts handle those."
4. Launch experts in parallel. Dispatch all expert prompts simultaneously. Set a uniform timeout (default: 120 seconds). Record launch timestamps.
5. Collect and normalize outputs. When each expert returns, extract:
   • Their recommendation (a single concrete action or choice).
   • Their confidence level (HIGH / MEDIUM / LOW).
   • Their top 3 supporting arguments.
   • Any risks or caveats they identified. Normalize into a uniform table: | Expert | Recommendation | Confidence | Key Arguments | Risks |.
6. Detect consensus. Count how many experts agree on the same recommendation. Apply rules:
   • All agree → STRONG_CONSENSUS.
   • Majority agrees → MAJORITY_CONSENSUS.
   • No majority → DISAGREEMENT.
7. Resolve disagreements. If DISAGREEMENT, identify the specific axes of conflict. List each contested point with the arguments for and against from each expert. If possible, run a brief tiebreaker prompt that presents only the contested points.
8. Synthesize the final recommendation. Write a single recommendation that:
   • States the chosen action and why.
   • Incorporates the strongest risk mitigations from dissenting experts.
   • Assigns a final confidence level based on consensus strength.
9. Document the panel record. Preserve the full expert outputs, the synthesis reasoning, and the final recommendation for audit purposes.

Decision rules

• Never average opinions — pick a concrete recommendation and defend it with evidence.
• Dissenting expert views must be acknowledged, not silently discarded.
• If all experts are LOW confidence, escalate to a human rather than synthesizing a weak answer.
• Prefer the recommendation with the fewest unmitigated high-severity risks.
• Weight domain-specialist opinions higher than generalist opinions on domain-specific questions.

Output requirements

1. Panel Composition — table of experts, their specialties, and perspective focus.
2. Expert Outputs — normalized table with recommendations, confidence, arguments, and risks.
3. Consensus Assessment — STRONG_CONSENSUS, MAJORITY_CONSENSUS, or DISAGREEMENT with details.
4. Synthesis — final recommendation with rationale and incorporated risk mitigations.
5. Dissent Record — any dissenting views and why they were not adopted.

References

• references/delegate-contracts.md — prompt templates for expert agents.
• references/checkpoint-rules.md — when to checkpoint before acting on a panel recommendation.
• references/failure-escalation.md — escalation triggers when consensus is unachievable.

Related skills

• parallel-lane-safety — for ensuring expert agents do not interfere with each other.
• manager-hierarchy-design — for structuring teams when experts become ongoing workers.
• verification-before-advance — for validating the synthesized recommendation before execution.
• long-run-watchdog — for monitoring expert agents that take too long.

Failure handling

• If an expert times out, proceed with the remaining experts but note reduced panel size in the synthesis and lower the overall confidence by one level.
• If all experts produce identical boilerplate, the prompts lacked differentiation — rewrite the role-specific instructions with sharper focus areas and re-run.
• If the synthesis contradicts all expert recommendations, discard the synthesis and present the raw expert outputs to the user for manual decision.
• If fewer than 2 experts return usable output, abort the panel and fall back to a single best-qualified agent with an enriched prompt.