Drill: delegate or do inline?

Delegate or do inline?

Interactive decision tree — outcomes:

• Do it inline

  The answer is smaller than the brief. Delegation would pay full fresh-context overhead and a spawn round trip to protect the parent from a fact that was never a threat.

• Keep it inline — manage the window instead

  A brief is written once and cannot answer questions. If the subtask needs mid-flight judgment, keep it where the conversation lives, and control context growth with summarization or compaction. Or restructure until the brief stands alone.

• Do it inline

  No disposable bulk, no risk boundary — delegation here adds an information-loss hop and overhead while buying nothing. The default is one agent until the context objects.

• Delegate: one subagent, tight brief

  Goal with success criteria, constraints including known findings, and a return format demanding evidence pointers and uncertainties. The bulk dies with the subagent’s window; the finding comes home.

• Fan out and synthesize

  One brief template, one variable slot, identical return formats. Spawn in parallel — wall-clock approaches the slowest worker — then synthesize in a clean window, spending targeted follow-up briefs on gaps and contradictions.

• Delegate to an isolated, credential-free subagent

  Narrow tools, no secrets, sandboxed if possible — enforced by the runtime, not the prompt. The untrusted material goes down; only a verdict comes back up. You have removed a lethal-trifecta leg structurally.

• Stop — restructure before running this

  Risky work that cannot be briefed to stand alone should not run inline with full privileges either. Split the task differently, add human approval gates, or narrow what the risky step can touch — the shape of the task is the problem.