Amey-Thakur Subagent delegation
Delegating to subagents buys parallelism and context isolation at the cost of coordination. It helps when parts are genuinely independent and hurts when the work needs one coherent view.
Method
- Delegate genuinely independent work. Searching several areas, reviewing along different dimensions, or handling separate files (see agent-map-reduce).
- Give each subagent a complete brief. It cannot see your conversation, so the task, constraints, and output format must be self-contained (see agent-handoff-protocol).
- Specify the return shape. Structured results are combinable; free-form prose from five subagents is another synthesis problem (see output-format-control).
- Do not delegate work needing continuity. Sequential reasoning where each step depends on the last belongs in one context.
- Reserve the main context for synthesis. Subagents gather and filter; the orchestrator decides, which is what keeps the decision coherent.
- Cap the fan-out. Each subagent costs tokens and time, and beyond a handful the coordination exceeds the benefit.
- Verify important subagent output. Delegated work is unreviewed by default, and confident wrong results propagate into the synthesis (see agent-generate-and-verify).
Boundaries
Delegation multiplies cost, so it needs justification beyond appearing thorough. Subagents cannot ask clarifying questions mid-task. Results are only as good as the brief, and vague briefs return vague work.