Subagent
A subagent is a separate agent instance, with its own context window, system prompt and tool access, to which a coding agent hands a task and from which it receives only a summary of the result.
The point is isolation. Whatever the subagent reads (the files it opens, the logs it scans, the search results it throws away) stays in its own context, and the parent conversation receives the summary and nothing else. In Claude Code, subagents are Markdown files with YAML frontmatter in .claude/agents/ or ~/.claude/agents/, and Explore, Plan and a general-purpose agent come built in (as of September 2026).
That isolation is where the saving comes from, and the only place. The main conversation stays small, so every later turn in it re-sends less. Claude Code's cost guidance recommends delegating verbose work, such as running tests, fetching documentation or processing logs, for exactly this reason.
The costs are real as well. A subagent starts without the parent's history and may have to gather context the parent already had. Its tokens still count: Claude Code's /usage screen attributes part of a plan's usage to subagents. It also keeps its own prompt cache, which does not read the parent's and, as of September 2026, lasts five minutes by default on a subscription rather than the main conversation's hour.
The trade therefore favours tasks that read a lot and report a little. A delegated task whose result is as long as its input saves nothing, and giving a subagent a cheaper model, model: haiku in its Claude Code definition, lowers the price of all the reading it does.