When To Use Subagents

Subagents are useful only when there is a real boundary or leverage point.

Good Reasons

Use subagents when there is:

• specialist knowledge boundary
• different tool or permission boundary
• reusable service-like capability
• parallelizable independent work
• a large task that would pollute the parent context
• maker-checker validation need
• long-running background work
• an isolated implementation or research branch

Bad Reasons

Avoid subagents when:

• one agent can do the work cleanly
• the subtask is tiny
• the overhead outweighs the benefit
• the split would add latency or confusion without quality gains

Rule

Do not create a separate agent for every subtask.

Prefer single-agent execution unless delegation clearly improves quality, speed, isolation, or reliability.

Classify the next step before delegating:

• if the parent is blocked on the result immediately, the task is often better handled locally
• if the task can run in parallel without blocking the next local step, it is a better subagent candidate