orchestrate
Multi-Agent Orchestration
Meta-orchestration patterns mined from 597+ real agent dispatches across production codebases. The skill maps orchestration strategy to work shape, brief structure to agent type, and background/foreground to supervision and integration needs.
Core principle: Match the strategy to the work, partition agents by independence, inject enough context that parallelism is real, and let review overhead adapt as trust earns itself. The strategies below are reference patterns. Pick the one that fits, blend two when the work is mixed, invent your own when the patterns don't match.
Dispatch Surface by Host
The strategies are host-agnostic; the fan-out verb differs:
| Host | Fan-out surface | Notes |
|---|---|---|
| Claude Code | Agent tool — parallel calls in one block, background for swarms |
Worktree isolation via the agent's isolation option |
| Codex | spawn_agent with a role-appropriate agent_type |
Delegation gate precedence below |
| Pi (pi-nova pack) | dispatch tool with "mode": "parallel" |
Children inherit the safety gate via PI_CODING_AGENT_DIR |
| Any host, no fan-out verb | Task queue as bus — Sibyl tasks + worktree isolation | A real dispatch mode, not a degraded one |
Codex delegation gate precedence: contract-mandated verification counts as warranted subagent work — or route it through external CLI review processes, which don't count as subagents. Exploratory swarms still need the user's ask. Standing user grants ("spawn subagents any time you want") persist across sessions; record them in memory.