What Are Subagents?

A subagent is its own agent instance with its own context, sent out by a main agent to handle a subtask.

Core idea

Instead of doing a subtask itself, a main agent can delegate it to a subagent - a separate agent instance with its own, usually empty context, that gets exactly one task and reports back a result.

What subagents are good for

Parallel search: multiple subagents can search different parts of a codebase or different sources at the same time
Specialization: a subagent can be tailored to one specific job (e.g. only code review, only research)
Context hygiene: the subagent accumulates many intermediate steps in its own context - the main agent only gets the final result back, instead of the whole research trail filling its context window

The cost tradeoff

Subagents cost extra tokens - each one has its own startup and its own intermediate steps. Multiple parallel subagents can noticeably raise the cost of a run. The benefit (a clean main context, parallel work) has to justify that.

Example: Opus 4.8

Claude Opus 4.8 introduced "Dynamic Workflows" as a research preview in May 2026: Claude Code can orchestrate hundreds of parallel subagents - an example of how far subagent concepts have been pushed in current agent tools.