Subagents run an independent task in their own context window and report back a conclusion. They’re ideal for broad searches, multi-file audits, and anything you’d otherwise grind through serially.

Why it matters

Two wins:

Parallelism. Three explorers finish in the time of one. A task that decomposes into independent pieces should run as several agents at once, not one after another.
Context hygiene. The file dumps and dead ends stay inside the subagent. Your main session keeps only the conclusion, so it doesn’t bloat and trigger context compaction.

How to improve

When a task splits into independent pieces, launch one subagent per piece in a single message so they run concurrently.
Use subagents for “go find out” work — searching a large codebase, auditing across many files, gathering options — and keep the synthesis in your main session.
Give each subagent a crisp, self-contained brief and ask it to return a conclusion, not a transcript.