subagents

Anthropic's Claude Code documentation, published 2025-09-29, defines subagents as delegated agents that each run in their own separate context window, with a configurable tool set and a custom system prompt of their own. The delegation is what matters here: the main conversation stays on the high-level objective while the intermediate output stays somewhere else.

Thariq, an engineer at Anthropic, frames subagents as one of five options a user has at the end of a turn, alongside continuing, rewind, `/clear` and `/compact`. His test for when to reach for one is a question about the output rather than the task: will I need this output, or only the conclusion? Verifying a result, reading another codebase, writing documentation. All of them generate volume that the main session never has to carry.

The second use is adversarial. «Boris Cherny» spawns several subagents for code review that challenge each other: a first pass on style guidelines, project history and obvious bugs, then five more poking holes in the original findings. At Every, the assess step of «Compound Engineering» runs twelve parallel subagents scoring the same code on security, performance, complexity and architecture, and Every has released that workflow as an open source plugin for «Claude Code».

The documentation lists chaining several subagents for complex workflows, and dynamic selection where Claude picks the appropriate one from the task description. Thariq expects these mechanisms to eventually be handled automatically. For now, deciding when to delegate remains the user's judgement.