Subagents

Quick Take

In brief: A subagent is an agent delegated a bounded part of a larger task.
Useful for: Separate research, review, or implementation work.
Main limit: Isolation, tools, lifetime, and edit access depend on the host.

Overview

A parent task can delegate an independent test review to a subagent and continue investigating code. The parent then checks the report before using it.

Core Definition

A subagent is an agent instance or process assigned a subtask by another agent or orchestrator. The host defines how it receives context, which tools it can use, whether it can edit files, and how its result returns.

Background and Evolution

Agent frameworks use delegation to split larger tasks. Some run subtasks in parallel; others use a sequential handoff. The term does not imply a specific hierarchy or execution environment.

How It Works

The parent or orchestrator defines a subtask and expected output.
The host starts a subagent with selected context and permissions.
The subagent completes work or reports a blocker.
The parent reviews and integrates the result.
The parent verifies the combined task.

Applications in Software Development and Testing

Delegate independent research, a review of a specific directory, or a test audit. Avoid overlapping file edits unless the host provides a safe coordination method.

Strengths and Limitations

Delegation can reduce the time needed for independent work. It adds coordination and review. A subagent may use incomplete context, return conflicting advice, or fail to finish.

Comparison with Similar Terms

A subagent is defined by delegation within a larger task. A background agent runs outside the immediate interaction. Multi-agent orchestration coordinates multiple agents. Product implementations may combine these patterns.

Best Practices

Set a narrow scope, a baseline revision, and a clear deliverable. Explain whether edits are allowed. Review the output and rerun relevant checks before integration.

Common Pitfalls

Do not assume a subagent is isolated or has limited permissions. Do not accept its findings without checking evidence. Do not delegate tightly coupled edits to multiple agents without coordination.

FAQ

Do subagents always work in separate branches?

No. The host decides where they run and whether their changes are isolated.