delegate

Delegate

Part of the Build phase — see the workflow skill for tiers and sequencing.

Subagents exist to protect the main context window and exploit true parallelism — not to add process. Every dispatch costs a full agent spin-up: the agent re-reads context, explores, works, and reports. Spend that cost only when it buys something.

Triggers — delegate when (and only when)

1. Context-flooding search/read: answering requires sweeping many files but you only need the conclusion, not the file contents.
2. Verbose output: test suites, builds, long logs — where only the summary matters in the main session.
3. True parallelism: 2+ genuinely independent tasks with no shared state and no sequential dependency.
4. Isolated large workstream (T3): big enough that a fresh, focused context outperforms a crowded one.

When a trigger fires, stop and call AskUserQuestion: name the task, why it triggered, a suggested model tier, and delegate-or-inline as options. Never decide silently in either direction. This applies at the trigger point, whenever it fires — including in the middle of another skill's process. (Doesn't apply to the anti-triggers below.)

Anti-triggers — do it inline instead:

• Small, sequential, or interdependent work.
• A 30-line change never needs an agent round-trip.
• Never delegate "to follow process" — an implementer+reviewer pair per tiny task turns 15 minutes of work into 8 round-trips of ceremony.
• Exploratory debugging where you don't yet know what's broken — you need the full picture in one context.
• Related failures (fixing one might fix the others): investigate together first.

Writing the dispatch prompt

A subagent inherits none of your session. You construct exactly what it needs:

1. Focused scope — one clear problem domain. "Fix all the tests" loses the agent; "fix agent-tool-abort.test.ts" keeps it on target.
2. Self-contained context — paste the error messages, name the files, one line on where this fits in the project, interfaces/decisions it can't discover itself, and your resolution of any known ambiguity.
3. Constraints — "do NOT change production code", "fix tests only", "don't just increase timeouts — find the real issue."
4. Expected return — say exactly what to report back: "return a summary of root cause and changes", a structured result, a file path. The agent's final message is your only deliverable from it.
5. Exact values live in one place. If the work follows a plan, point the agent at its task text rather than paraphrasing requirements — paraphrase drift is how implementations diverge from spec.