Sub-agent

A sub-agent is an agent that handles a delegated part of a larger task for a parent agent, usually with its own instructions, its own tools and a separate context window.

Why use one. Anthropic’s research system runs sub-agents in parallel, each exploring one angle of a question with a clean context and returning condensed findings to the lead agent. Its guidance on context engineering (September 2025) gives the same reason in general terms: detailed exploration stays inside the sub-agent, and the parent receives a short summary, which keeps the parent’s context window focused. Giving each sub-agent only the tools it needs also limits what any one of them can touch.

In frameworks.

• Google ADK: agents listed in a parent’s sub_agents form a hierarchy. In collaborative workflows (ADK 2.0 and later), ADK generates a delegation tool for each sub-agent, named after it, and control returns to the coordinator when the sub-agent finishes.
• Scoped tools: ADK’s documentation shows wrapping a sub-agent with AgentTool and giving it its own MCP toolset, narrowed with tool_filter to the few tools that sub-agent needs.
• Remote sub-agents: ADK’s RemoteA2aAgent lets a remote A2A agent, located through its Agent Card, sit in the tree like a local sub-agent.

Local and remote. A local sub-agent shares the parent’s process, credentials and trust boundary. A remote one, reached over A2A, is a separate system run by someone else. The parent sees only its Agent Card, the task states and the returned artifacts, and should treat that output as untrusted input, like any other external content, before acting on it.

Neighbouring terms. The orchestrator agent is the parent role. A multi-agent system is the whole arrangement. A remote agent is the A2A term for the agent that receives a task.