Delegate or Die (Protect the Main Thread)

The main conversation is for decisions, not for bulk reading.

Every grep result, file dump, and fetched page that lands in the main context is re-paid on every later turn, crowds out the original goal, and accelerates context rot. The agent that keeps its main thread clean is the agent that still knows what it's doing an hour in.

1. When to delegate

Delegate when any of these is true:

• The answer requires reading more than ~3 files, or enters unfamiliar territory.
• The task is a search sweep: multiple queries, page fetches, or comparing options.
• There are 2+ independent tasks with no shared state — spawn parallel subagents.

Handle inline only:

• Single-file edits
• Ambiguous scope where you need the user in the loop
• Destructive operations
• Purely mechanical changes (renames spanning a few files)

When torn between two sizes, pick the smaller delegation.

2. The brief

Subagents know nothing. They inherit no conversation, no project memory, no intent. Every brief must be self-contained:

1. The exact question — one sentence, not a topic.
2. Boundaries — read-only or allowed to write; which paths are in scope.
3. Relevant file paths and constraints — anything they'd waste calls discovering.
4. The output contract — what the final message must contain and must not contain.

Quality over budget. Do not cap tokens or tool calls — caps make agents timid and produce truncated work. Define the output contract instead; contracts make agents precise. The goal is one-pass completion, not cheapest completion.

Adversarial brief (doubt review). The refuter variant strips the brief to two parts: the ARTIFACT (the decision or plan to attack, with zero author justification) and the CONTRACT (acceptance criteria). The mission is only to refute — no validation, no summaries; every gap reported as claim → why it fails → what evidence would change your mind. Cap at 3 cycles, then escalate to the human. Full procedure: the doubt-driven-development skill.

3. The output contract

Every delegated task returns:

• Status line first, before any content:
  • STATUS: DONE
  • STATUS: DONE_WITH_CONCERNS — name the doubts
  • STATUS: BLOCKED — name the blocker
  • STATUS: NEEDS_CONTEXT — name what is missing
• A structured synthesis with file paths and line ranges, so claims are checkable.
• Never raw dumps, whole-file pastes, or transcripts of the search process.

A BLOCKED or NEEDS_CONTEXT subagent returns early with just the status and a one-line reason — it must not pad the report to fill the format.

4. After the return

• Don't repeat the delegated work in the main thread. Once dispatched, wait for the result — re-running the search yourself doubles the cost and re-floods the context you just protected.
• Verify before acting. Check cited paths and line ranges before building on a subagent's claims; a synthesis is cheaper to verify than to redo.
• Relay what matters. The subagent's report is not shown to the user — surface the conclusion in your own words.