Subagents

One at a time. Tell it whether you knew, and the ones you did not stay in your pile.

1 / 8

A subagent has its own context window and returns only a summary.

A big read stays out of your window. It buys parallelism or fresh eyes, not a smarter answer.

The whole topic8

A subagent has its own context window and returns only a summary.
A big read stays out of your window. It buys parallelism or fresh eyes, not a smarter answer.
The built-in Explore agent skips CLAUDE.md and git status.
It is read-only search, kept light on purpose, so do not expect it to know your project rules.
isolation: worktree gives a subagent its own temporary git worktree.
It is removed automatically if nothing changed, so risky edits do not touch your checkout.
A subagent's tools list is enforced by the harness, not requested in a prompt.
tools: Read, Grep, Glob makes a reviewer that cannot edit anything.
A subagent's model resolves per-call, then frontmatter, then CLAUDE_CODE_SUBAGENT_MODEL, then the main model.
So a cheap model: haiku in the file is enough, unless a call overrides it.
Misspelled subagent frontmatter fields are ignored silently.
Run claude --debug and search for "does not recognize or apply" to find them.
A fork inherits the parent's history exactly, so it reads the parent's cache.
A normal subagent starts cold. Use /fork when the side task needs the whole conversation.
@agent guarantees a subagent runs; --agent runs the whole session as one.
Without a mention, Claude decides on its own whether to delegate.