Subagent

Subagent

Also called: 子智能体 · Sub-agent

An agent spawned by a parent agent to handle one bounded subtask, with its own context window and tool permissions, returning a result to the parent.

Main agentSubagent: searchSubagent: testSubagent: summarise
Schematic (simplified)

Why it exists

Two main motivations:

  1. Protect the parent context — hand work that reads in a lot of intermediate material (crawling a codebase, scanning a dozen web pages) to a subagent, so the parent receives only a condensed conclusion instead of a context full of noise.
  1. Isolation and specialisation — give the subagent its own system prompt and a narrower tool set so it stays focused on one job, and its mistakes do not pollute the main flow.

Difference from a plain tool call

A tool call is one deterministic function execution. A subagent is a full agent loop — it reasons, calls tools multiple times, self-corrects, and only then returns. Think of it as "a tool you can outsource a whole task to."

Common misconceptions

  • A subagent does not share the parent's context; it sees only the input the parent explicitly passes it.
  • More subagents is not necessarily faster: spawning, waiting, and merging all cost time.

Related terms

Sources

  1. Anthropic — Claude Docs: Subagents
  2. Anthropic — How we built our multi-agent research system (2025)

Compiled 2026-08-29 · This glossary is compiled from public papers, official specifications, and common industry definitions, and is updated as the field evolves. Corrections welcome.