Multi-Agent Orchestration

Do not add agents. Add independently verifiable work: a dependency graph, parallel only on independent nodes, integrate through shared state and a verifier.

When to use

• About to spawn 2+ subagents, a research fan-out, a debate/jury, or a "programmer + reviewer" pair
• Designing a coding-agent DAG, worktree split, or long-running parallel implementation
• Tempted to scale agent count because the task "feels hard"

When not to use:

• One tightly sequential transform (read → patch → test) with no independent parts
• Same model, same input, same tools, only the role name changes — run multiple turns on one agent instead

Gate

Default to a single agent. Go multi only if (1) or (2) holds and (4) is positive.