Cursor subagents are specialized agents with their own context. They can run in the editor, the command line interface, and Cloud Agents.

The value is isolation. A subagent can inspect one concern deeply without polluting the main agent’s context with every intermediate detail.

Built-In Subagents

Cursor includes built-in subagents such as Explore, Bash, and Browser. They are useful for bounded tasks:

• Explore: inspect a repository or design space.
• Bash: run terminal-heavy investigation.
• Browser: inspect web or application behavior.

Use them when the task shape matches the tool. Do not split work just to make the prompt look sophisticated.

Custom Subagents

Custom subagents live in .cursor/agents/:

---
name: test-auditor
description: Find missing regression coverage for a described bug.
model: auto
readonly: true
---

Inspect the changed files and closest tests. Report missing coverage with file
references. Do not edit files.

Cursor also supports compatible agent definitions from .claude/agents and .codex/agents, which is useful in mixed-agent repositories.

Foreground and Background Work

A foreground subagent blocks the main task until it returns. A background subagent can continue while the main agent proceeds. Background work is useful for independent research, but it needs a crisp return format:

Return only: findings, file references, and whether the main task should stop.

Without a return contract, parallel work often creates more context than value.

When Not to Use a Subagent

Do not use subagents for tiny edits, single-file fixes, or tasks where one developer needs to keep the whole mental model in view. Subagents are a way to reduce context pressure, not a way to avoid thinking.