Delegate-check

Before spawning a sub-agent, answer one question out loud:

Will I need this tool output again (raw file contents, full grep results), or just the conclusion?

• "Just the conclusion" → delegate. The subagent burns its own context window; the main thread stays planning-clean.
• "Raw output" → do it yourself. The subagent's summary will strip exactly what you need to keep.

Output contract (required for every Agent invocation)

A vague handoff produces a dumping-ground report that re-pollutes the main context, defeating the whole point. State three fields explicitly in the prompt:

1. Goal — one sentence. What single question does the agent answer?
2. Allowed tools — minimum tool set. Restrict explicitly. Examples:
   • "Read, Grep, Glob only. No Bash, no Edit, no Write."
   • "WebSearch + WebFetch only. No file tools."
3. Return format — what the last message must contain, with explicit do-nots:
   • "List of file paths, one per line. Do not return raw code or excerpts."
   • "Top 3 candidates, one-line trade-off each, ≤200 words total. No prose narration of process."
   • "Pass/fail verdict + 3 specific findings. No 'observations' section."

Parallel sub-agents — the two-humans test

Spawn parallel subagents only if you would give the same tasks to two humans without them needing to talk first:

• 3+ independent tasks
• No shared mutable state
• Clear file / area / topic boundaries

If any of those fails, run sequentially. Parallelism without independence causes race conditions, duplicate edits, and stale reads.

When NOT to delegate

• Self-reflection ("what did I just do?", "what should I do next?") — introspect in main context. Spawning for this is a documented failure mode.
• Decisions you must make — the subagent will give you a summary, you still decide. If decision is the whole task, you cannot delegate it.
• Small lookups (<5 reads, single grep) — direct tool use is cheaper than spawn-overhead.
• Tasks where you need the raw output — delegation strips detail.

Anti-patterns

• ❌ Spawning without specifying return-format → 800-word narration polluting main context.
• ❌ Granting * tools when read-only would suffice → subagent burns its own context on irrelevant tool docs.
• ❌ Parallel subagents on tasks that share files / state → race conditions.
• ❌ Forked subagent (same context) for review work → anchoring bias. Use fresh context for review.
• ❌ Subagent for "think harder" — the subagent thinks no harder than the parent. Use it for exploration, not cognition.

$ARGUMENTS