TL;DR
A specialized, auxiliary AI agent delegated by a parent or orchestrator agent to complete a highly specific, bounded task within an independent context window.
In modular agentic architectures, a parent agent manages the overall objective and calls subagents to solve specific technical steps, such as codebase exploration, local execution, or deep web searches. Subagents run their own separate multi-step reasoning loops, utilize targeted toolsets, and operate in isolated context windows. Once their work is completed, they pass a condensed summary back to the parent agent, preventing the main conversation context from being bloated with intermediate logs and step-by-step outputs.
Why this matters for your business
Subagents reduce token costs and memory overhead in complex workflows while allowing developers to implement highly parallelized, domain-specific execution loops within larger autonomous systems.
