agents

A collection of reusable AI subagents for Cursor, organized by category. Each agent is a single Markdown file with YAML frontmatter that Cursor's Agent can delegate to automatically or you can invoke explicitly with /name.

The design goal: small, focused, single-purpose agents that compose. Project-specific rules (stack, commands, conventions) live in each project's AGENTS.md, so these agents stay generic and portable.

Categories

Agents are grouped by domain. Add new categories as folders; each category has its own README.md.

Category	What lives here	Status
dev/	Building and maintaining code — orchestration, review, simplification, testing, docs	✅ populated
ops/	Infrastructure, deployment, CI/CD, incident response	🌱 placeholder
research/	Codebase exploration, dependency & security audits, technical spikes	🌱 placeholder
writing/	Non-code writing — changelogs, release notes, PR descriptions, announcements	🌱 placeholder

Install into a project

Cursor loads subagents from .cursor/agents/ (project) or ~/.cursor/agents/ (all projects). Copy the agents you want from a category into that folder:

# from your project root — pull in the whole dev fleet
mkdir -p .cursor/agents
cp /path/to/agents/dev/*.md .cursor/agents/

Or symlink a category so updates flow through:

ln -s /path/to/agents/dev .cursor/agents

Cursor also reads .claude/agents/ and .codex/agents/ for cross-tool compatibility, so the same files work there.

Commit .cursor/agents/ into each project so your team gets the agents too.

How the dev fleet works together

The dev/ category is a coordinated fleet, not five unrelated tools. lead orchestrates the rest:

plan → build → test → simplify → review → docs

See dev/README.md for the full pipeline, parallel patterns, and the shared auto-fix safety policy.

Conventions for agents in this repo

• One responsibility per agent. No generic "helper" agents.
• Invest in the description. It's what Cursor reads to decide when to delegate — be specific about when to use the agent.
• Read AGENTS.md first. Every agent should defer to the host project's conventions over its own defaults.
• Keep prompts concise. Aim for well under ~200 lines; long prompts dilute focus.
• Declare write scope. Reviewers/auditors set readonly: true. Writers follow an explicit auto-fix policy (apply low-risk, flag risky).

Adding a new agent

1. Pick or create a category folder.
2. Add <name>.md with frontmatter (name, description, model: inherit, and readonly: true for non-editing agents).
3. Write a focused prompt: when invoked → what to check → what to output.
4. Update that category's README.md table.

Adding a new category

Create a folder, add a README.md describing its purpose and listing its agents, then add a row to the table above.

License

MIT (or your choice — add a LICENSE file before publishing).