Subagent

A subagent is an AI agent invoked by a parent or orchestrator agent to execute a specific bounded subtask autonomously and return results, without direct interaction with the end user during its execution.

A subagent is a specialized or general-purpose AI agent that is invoked by another agent (the orchestrator) rather than directly by a human user. The orchestrator delegates a bounded subtask—such as web search, code execution, document summarization, or API interaction—and the subagent completes it within its own context window and tool set before returning a structured result.

Subagents typically run as isolated processes or API calls. Each subagent receives a focused prompt containing only the context relevant to its subtask, reducing noise and preserving context budget. Some architectures give subagents their own persistent memory or tool access; others keep them stateless and ephemeral. The parent agent synthesizes results from one or more subagents to continue reasoning or produce a final output.

The subagent pattern is valuable because it enables parallelism—multiple subagents can execute simultaneously on independent subtasks—and specialization, where different subagent instances use different models, prompts, or tool sets optimized for their domain. It also improves fault isolation: if one subagent fails, the orchestrator can retry or reroute without restarting the entire pipeline.

In the Claude ecosystem as of 2026, subagents are a first-class concept in the Claude Agent SDK, where a parent Claude instance can spin up child agents using the Task primitive. Similar patterns appear in AutoGen, CrewAI, and LangGraph, and are used in production systems for automated code review, multi-step data analysis, and document processing at scale.