| name | multi-agent-deliberation |
| description | Use for high-stakes decisions: architecture forks, build-vs-buy, competitive response, major refactors, or when the user wants multi-agent debate. Spawns diverse perspectives (or Codex second opinion) and synthesizes a recommendation. Do NOT use for trivial choices or every small implementation detail.
|
Multi-Agent Deliberation
When a wrong call is expensive, prefer competing lenses — then decide. Optional: skip for clear choices or when the user wants speed.
When to Use
- Build vs buy / build vs wrap competitor
- LangGraph vs Pi vs custom orchestration (after
agent-framework-choice still unclear)
- "Should we pivot given competitor X?"
- Large rewrite vs incremental
- User explicitly asks for multi-agent or second opinion
Procedure
1. Frame the decision
Write:
- Question (one sentence)
- Constraints (time, skills, stack, distribution)
- Success metric
- Deadline / reversibility (one-way door?)
2. Run 2–3 lenses (parallel when possible)
Pick lenses that disagree productively:
| Lens | Focus |
|---|
| User/value | Does this help the buyer this month? |
| Risk/security | What fails in prod? |
| Tech debt / YAGNI | Simplest path (ponytail spirit) |
| Competitive | How does this position vs named alternatives? |
| Codex/rescue | Independent second implementation or review |
Implementation options:
- Claude Code
Agent tool with distinct system prompts per lens
- Codex plugin
/rescue or /adversarial-review when available
- Sequential self-roles only if subagents unavailable (label each role clearly)
3. Synthesize (you, main agent)
Produce:
## Decision question
## Options
## Lens votes (table)
## Recommendation
## Dissent / residual risks
## What would change the decision
## Proposed next experiment (smallest test)
Do not average into mush. Pick a recommendation. Surface dissent.
4. User owns the call
Present synthesis; user confirms. Record in docs/decisions/ or dev-log when multi-session.
Anti-Patterns
- Five agents agreeing with the same prompt (fake diversity)
- Deliberating UI padding color
- Hiding the recommendation behind "it depends" with no default
- Skipping research gates by debating vibes only — attach evidence
Related
product-bootstrap, research-before-build, agent-framework-choice