Subagent

A subagent is an agent launched by another agent to handle a bounded piece of work. It runs with its own context window, often its own system prompt and tool list, does the job, and returns a short report. The parent never sees the subagent's intermediate steps: the twenty file reads, the failed searches, the log dumps all stay in the child's window and are discarded when it finishes. Only the conclusion crosses back.

This is mostly a context engineering tool. A search across a large repo can pull thousands of lines into context; done in a subagent, the parent pays for a paragraph. It also enables parallelism, with several subagents working different parts of a task at once. The costs are real too: each subagent starts from zero, so it needs a self-contained brief, and it cannot see what the parent has already learned unless the parent tells it.

Delegate work whose output is a conclusion, not a conversation: research, broad searches, independent reviews, isolated implementation in a worktree. Write the brief as if for a new hire with no history, including what done looks like. Keep tight, sequential edits in the main session where the state lives. Verify what comes back: a subagent can be confidently wrong, and the parent has no transcript to check, only the report.