slgoodrich Team Coordination
Protocols for structuring productive debate, cross-examination, and synthesis across multi-agent teams.
When to Use This Skill
Auto-loaded by all six agents:
idea-researcher,market-researcher,idea-skeptic- For structured debate and cross-examination in validation sprints and war roomsmarket-fit-reviewer,feasibility-reviewer,scope-reviewer- For cross-referencing in PRD stress tests
Use when you need:
- Running multi-agent team workflows
- Structuring productive debate between agents
- Cross-examining findings from parallel investigations
- Synthesizing multiple perspectives into a verdict
- Handling disagreements between team members
- Building consensus or documenting dissent
Core Principle: Structured Disagreement Produces Better Decisions
Solo decision-making has blind spots. But unstructured debate is just noise. The sweet spot is structured disagreement: agents with different perspectives challenge each other using evidence, following clear protocols that force productive conflict.
Not this: Three agents agreeing with each other to be polite. This: Three agents stress-testing each other's conclusions with pointed questions and evidence.
The goal isn't agreement. It's a well-examined conclusion where every weakness has been probed.
Debate Protocol
Phase Structure
Every team debate follows three phases:
1. Independent Investigation Each agent works in parallel on their assigned dimension. No cross-talk. This prevents groupthink and ensures genuinely independent perspectives.
- Each agent produces a structured finding report
- Reports include evidence, confidence levels, and key claims
- No agent sees another's work until investigation completes
2. Cross-Examination Agents receive each other's findings and challenge them. One round of structured challenges. Each agent must:
- Cite specific claims from other agents' reports
- Challenge with evidence, not opinion
- Ask pointed questions that could change the conclusion
- Acknowledge when another agent's evidence is strong
3. Synthesis The lead agent compiles all perspectives into a final deliverable. Must:
- Represent all perspectives accurately
- Note where agents agreed (convergence)
- Note where agents disagreed (divergence) and why
- Make a final recommendation with explicit reasoning
- Flag unresolved tensions rather than papering over them
Debate Rules
Evidence over assertion. "I think this won't work" is weak. "Users in similar markets showed 3% conversion rates [source], suggesting this won't work" is strong.
Steel-man before attacking. Before challenging a claim, restate it in its strongest form. This proves you understood it.
One round, make it count. Cross-examination is a single round. Agents can't go back and forth endlessly. This forces precision.
Disagree and commit. After synthesis, the verdict stands even if an agent disagrees. But dissent gets documented, not silenced.
No false consensus. If agents genuinely disagree, the synthesis says so. "Researchers disagreed on X. Here's why." is better than pretending everyone agreed.
Cross-Examination Patterns
Challenge Formats
When examining another agent's findings, use these question structures:
Evidence Challenge
"You claim [X]. What evidence supports this beyond [the single source you cited]? If that source is wrong, does your conclusion still hold?"
Assumption Challenge
"Your analysis assumes [Y]. What happens to your conclusion if [Y] is false? Is there evidence that [Y] might not hold?"
Magnitude Challenge
"You identified [risk/opportunity]. How significant is this really? Is this a deal-breaker or a minor concern? What's the actual impact?"
Alternative Explanation
"You attribute [outcome] to [cause]. Could [alternative cause] explain this equally well? Have you considered [different framing]?"
So-What Challenge
"You found [data point]. What does this actually mean for the decision we're making? How should this change our recommendation?"
Challenge Quality Standards
Strong challenge (changes the conversation):
- Cites specific evidence that contradicts the claim
- Identifies an assumption the agent didn't examine
- Presents an alternative explanation with supporting data
- Asks a question the agent should have answered but didn't
Weak challenge (wastes a round):
- "I'm not sure about that" (no evidence)
- Repeating your own findings without engaging the other agent's
- Nitpicking minor points while ignoring major claims
- Asking rhetorical questions that don't demand real answers
Consensus vs. Dissent Handling
When Agents Agree
Document the convergence explicitly:
## Convergence Points
All three perspectives agree on:
1. [Finding] - Supported by [evidence from each agent]
2. [Finding] - Supported by [evidence from each agent]
Confidence: HIGH (independent analyses converged)
Agreement from independent investigation is a strong signal. It means the conclusion is robust across different analytical lenses.
When Agents Disagree
Not all disagreements need resolution. Some reflect genuine uncertainty.
Resolvable disagreements (one agent has stronger evidence):
## Resolved Disagreement: [Topic]
- Agent A claimed [X] based on [evidence]
- Agent B claimed [Y] based on [evidence]
- Resolution: Agent A's position is stronger because [reasoning]
- Agent B's concern remains valid as a risk to monitor
Unresolvable disagreements (genuine uncertainty):
## Unresolved: [Topic]
- Agent A: [Position] based on [evidence]
- Agent B: [Opposing position] based on [evidence]
- Why it's unresolved: [Explanation - insufficient data, different frameworks, etc.]
- Implication: This uncertainty should factor into risk assessment
- Recommended action: [How to resolve - more research, user testing, etc.]
Dissent Documentation
When the synthesis verdict goes against one agent's recommendation:
## Dissenting View: [Agent Name]
[Agent] recommends [alternative] because [reasoning].
This dissent is noted because [why it matters].
The majority verdict differs because [reasoning].
If [condition], revisit this dissent.
Synthesis Patterns
Weighted Synthesis
When combining perspectives with different relevance:
- Assign weight to each perspective based on the decision being made
- For idea validation: User Problem (40%), Market Opportunity (30%), Defensibility (30%)
- For PRD review: Market Fit (35%), Feasibility (35%), Scope (30%)
- Weights are guidelines, not formulas. Override when evidence demands it.
Narrative Synthesis
Don't just list findings. Tell the story:
- What we investigated (the question)
- What we found (key findings from each perspective)
- Where we agree (convergence points)
- Where we disagree (divergence points)
- What it means (interpretation and verdict)
- What to do next (recommended actions)
Verdict Framework
Every synthesis must end with a clear verdict:
- Strong conviction: "BUILD / DON'T BUILD / READY TO BUILD" with high confidence
- Conditional: "BUILD IF [specific conditions are met]" with conditions list
- Insufficient evidence: "NEEDS MORE EVIDENCE on [specific gaps]" with research plan
Never end with "it depends" without specifying what it depends on and how to resolve it.
Team Communication Standards
Message Structure
When agents send findings to teammates:
## [Agent Name] Findings: [Topic]
### Key Claims
1. [Claim] - Confidence: [HIGH/MEDIUM/LOW] - Evidence: [summary]
2. [Claim] - Confidence: [HIGH/MEDIUM/LOW] - Evidence: [summary]
### Evidence Base
- [Source 1]: [What it shows]
- [Source 2]: [What it shows]
### Gaps and Limitations
- [What I couldn't find or verify]
- [Assumptions I'm making]
### Preliminary Recommendation
[What I think this means for the decision]
Confidence Levels
- HIGH: Multiple independent sources confirm. Would bet on this.
- MEDIUM: Some evidence supports, but gaps exist. Directionally correct.
- LOW: Limited evidence, significant assumptions. Treat as hypothesis.
Common Anti-Patterns
1. Consensus Theater
Agents agree too quickly to avoid conflict. Every team debate should have at least one genuine challenge. If nobody disagrees, someone isn't doing their job.
2. Authority Bias
Deferring to the "lead" agent's view without genuine examination. The lead synthesizes; they don't dictate.
3. Evidence Arms Race
Agents competing to find the most evidence rather than the most relevant evidence. Five weak sources don't beat one strong one.
4. Scope Creep in Debate
Cross-examination expanding into entirely new investigations. Stick to challenging existing findings, not opening new lines of inquiry.
5. False Precision
Scoring with decimal precision (7.3/10) when the underlying evidence only supports rough buckets (high/medium/low). Don't manufacture certainty.
Ready-to-Use Resources
In references/:
- team-debate-patterns.md: Dialectical Inquiry, Devil's Advocacy, and Constructive Controversy frameworks with implementation details
Troubleshooting
"All agents agree and the debate produces nothing useful": Your prompts are too similar. Give each agent a distinct lens (optimist vs. skeptic, market vs. technical, user vs. business). Assign one agent the explicit devil's advocate role.
"The synthesis just averages the scores instead of finding insight": Averaging destroys signal. The synthesizer should identify where agents disagree most and explain why. A 5 and a 2 on the same dimension is more interesting than two 4s.
"One agent dominates the output": Check prompt length and specificity. If one agent's prompt is 3x longer, it will produce 3x more output and overwhelm the synthesis. Keep reviewer prompts roughly equal in scope.
Related
/agent-teams:validation-sprint- Uses these protocols for idea validation/agent-teams:prd-stress-test- Uses these protocols for PRD review/agent-teams:competitive-war-room- Uses these protocols for competitive analysis
Remember: The point of multi-agent debate isn't to be adversarial. It's to be thorough. Every blind spot found in debate is a blind spot that won't surprise you in the market.