subagents

# Subagents ## Purpose Delegate work to a subagent when its own context is worth the cost of starting one. A subagent begins cold — it re-derives everything the main agent already knows — so the delegation must buy more than it costs. ## When to Use - A search or exploration that would flood the main context with results you do not need to keep. - Independent work that can run in parallel. - A task requiring a different, specialized instruction set. - Verification by an agent that has not seen the reasoning that produced the work. ## Capabilities - Scoping a subagent's task and boundaries. - Prompt design for a cold-start agent. - Parallel fan-out for independent work. - Result aggregation. - Read-only agents for review and verification. ## Inputs - The task, and whether it is genuinely independent. - What the subagent needs to know, since it knows nothing. - What it must return. ## Outputs - A subagent that completes its task without further interaction. - A result compact enough to be worth the round trip. ## Workflow 1. **Justify the delegation** — The valid reasons are: context isolation (the search produces 50 files of noise and one answer), parallelism (five independent tasks), and specialization (a different instruction set). "The task is big" is not a reason — a big task in the same context is usually cheaper. 2. **Write the prompt as if to a stranger** — Because it is one. The subagent has none of your context: no conversation history, no prior findings, ...