| name | workflow |
| description | Multi-agent orchestration patterns for orchestrator agents. Three reusable patterns built on the existing call_subagent / wait_for_subagent / create_subagent / get_available_models tools. Available patterns: ensemble, pool, multi-model. |
Workflow Patterns
Category of orchestration patterns for composing existing subagent tools to solve common multi-agent problems. No new tools — every pattern is pure prompt-level guidance for the orchestrator.
Available Patterns
- ensemble: Run the same task on multiple agents/models in parallel and reduce the answers (vote / compare).
- pool: Run N tasks under a concurrency cap K (rate-limit-aware sliding window).
- multi-model: Recovery strategies when a subagent is stuck — fresh-context retry, or retry with a different model.
Usage
These patterns are read by orchestrator agents through their system prompt. Each sub-skill's SKILL.md describes when to use the pattern and lays out the recipe in pseudocode. They use only existing tools (call_subagent, wait_for_subagent, create_subagent, get_available_models, send_message, list_subagents) — no scripts.
Common Use Cases
- High-stakes question with multiple opinions needed (ensemble)
- Many independent subtasks under a shared rate limit (pool)
- A subagent has clearly stalled on a hard task (multi-model)
Requires Sandbox
None — pattern docs are pure prompt guidance.
