cursor-subagent-creator

Create a markdown file in `.cursor/agents/` (project-level) or `~/.cursor/agents/` (user-level) with YAML frontmatter defining name, description, model, readonly, and is_background fields. The body contains the subagent's instructions, steps, and expected output format.

Cursor subagents support five optional fields: name (identifier, defaults to filename), description (when to delegate), model (inherit, fast, or a specific model ID), readonly (restricts write permissions), and is_background (runs in background).

Use a subagent when the task is complex, has multiple steps, and benefits from isolated context, such as security audits or test failure investigations. Use a skill for single-purpose actions like formatting imports or generating a changelog.

Write a specific description stating when the subagent should be used, including phrases like "Use proactively when..." or "Always use for...". Vague descriptions like "helps with general tasks" give the Agent no delegation signal.

Yes. Multiple subagents can run simultaneously when the Agent sends multiple task calls in one message. Setting `is_background: true` makes a subagent return immediately and work independently, with state written to `~/.cursor/subagents/`.

Each subagent has its own context window, so parallel execution multiplies token usage and adds startup overhead. Dozens of vague subagents are ineffective; start with two or three focused ones and add more only for distinct use cases.