| name | swarm-orchestrator |
| description | Use when coordinating multiple AI agents for sports, fantasy, or market-analysis tasks, especially when the work benefits from role-based decomposition, disagreement analysis, critique loops, and a final synthesized recommendation across development or production runtimes. |
Swarm Orchestrator
Use this skill when one agent should not do all the reasoning alone.
Agent roles
Start with these roles unless the task clearly needs fewer:
- Research Scout
- Context Analyst
- Projection Analyst
- Risk Analyst
- Critic
- Final Synthesizer
Core loop
- Define the decision question precisely.
- Assign narrow subproblems to role agents.
- Require structured outputs from each agent.
- Compare outputs for agreement, disagreement, and missing evidence.
- Re-run only the weak or conflicting areas.
- Produce a final recommendation with assumptions and confidence.
Required output fields per agent
{
"thesis": "",
"evidence": [],
"assumptions": [],
"confidence": 0,
"known_unknowns": [],
"next_checks": []
}
Orchestration rules
- Do not ask every agent to do everything.
- Keep prompts role-specific and short.
- Prefer retrying one weak agent over rerunning the whole swarm.
- Treat disagreement as signal, not failure.
- Preserve the trace so outcomes can be evaluated later.
Synthesis rules
The final answer should always include:
- best current recommendation,
- main reasons,
- top risks,
- what would change the recommendation,
- confidence level,
- evidence gaps still open.
