✅ When a Sub-Agent Earns Its Complexity — and When It Doesn't

Sub-agents are easy to reach for once they're available, but they're not free — each one adds a coordination boundary that has to be designed and later audited. A useful rule of thumb: spin up a sub-agent when a piece of work is genuinely independent (its output doesn't need the parent's evolving context to make sense) and when keeping its exploration out of the main conversation protects context budget for the parts of the task that do need it.

• Good fit: "research how library X handles Y and summarise in 200 words" — self-contained, output-focused, doesn't need back-and-forth.
• Poor fit: anything that requires several rounds of clarification with you specifically — sub-agents can't ask you follow-up questions the way the main conversation can.
• Poor fit: work where you want to watch the reasoning unfold in real time — a sub-agent's intermediate steps aren't visible until it reports back.

✅ Naming and Scoping Sub-Agents So You Can Actually Audit Them Later

When a session has spawned several sub-agents, the difference between a five-minute audit and a confusing archaeology exercise usually comes down to naming discipline set up front. A sub-agent named descriptively for its actual task ("audit-auth-middleware", "migrate-user-schema") is trivially greppable later; a generic default name isn't.