Playbook

1. Decide Whether to Delegate

Delegate when the task is bounded, noisy, independently verifiable, or benefits from a specialist. Keep tightly coupled reasoning in the parent.

2. Assign Ownership

Name files, modules, or an analysis question. State that other agents may be changing the repository and that the child must preserve unrelated work.

3. Bound Authority

Specify read, write, command, network, and external-action permissions. Subagent creation does not broaden the user's original authorization.

4. Define the Result

Require a compact artifact: findings with evidence, a patch plus tests, or a decision table. Include uncertainty and unfinished work.

5. Isolate Writes

Use separate file ownership for disjoint changes. Use a worktree or environment when agents need overlapping paths or incompatible dependency state.

6. Verify in the Parent

Review the actual diff and rerun the load-bearing check after integration. A child's “tests passed” statement is a pointer, not proof.

7. Measure the Trade

Compare elapsed time, total model usage, defects, and integration effort with a single-agentAgentA worker you delegate to: brief it, and it takes steps on its own. A chatbot answers; an agent acts.An LLM that runs tools in a loop toward a goal — it acts, checks the result, decides the next step, and repeats until done.Full definitionClose baseline. More agents are justified only when the system outcome improves.