Skip to content

README.md

Latest commit

 

History

History
70 lines (52 loc) · 3.05 KB

README.md

File metadata and controls

70 lines (52 loc) · 3.05 KB

Agents glossary

What are agents?

Agents (subagents) are specialized AI assistants that run in their own context window. The parent agent delegates tasks to them automatically or you invoke them explicitly with /agent-name. They work in parallel, don't bloat your main conversation, and return a summary when done.

How they work

  • Each agent has its own context — long research or test runs don't eat your main conversation's memory
  • The parent agent reads the description field to decide when to delegate automatically
  • Agents can run in foreground (blocks until done) or background (returns immediately)
  • Agents can't launch other agents (single level only)

All agents

Agent Model Description
/verifier fast Validates completed work. Checks that implementations actually function, runs tests, catches edge cases. Use after tasks are marked done.
/security-reviewer inherit Security auditor. Checks for injection, auth bypass, hardcoded secrets, input validation, RLS gaps. Triggers proactively on auth, payments, API endpoints, or user data.
/test-runner fast Runs tests after code changes. Analyzes failures, fixes them while preserving test intent, reports results. Triggers proactively.
/researcher inherit Researches topics via web search and codebase analysis. Finds 5-7 sources, summarizes, flags contradictions. Read-only. Use during /explore or /critique.

How to invoke

Automatically — the parent agent delegates based on the task and agent descriptions. Phrases like "use proactively" in the description encourage this.

Explicitly — type the agent name as a slash command:

/verifier confirm the auth flow is complete
/researcher find competitors for X

Configuration fields

Field Required Description
name No Lowercase + hyphens. Defaults to filename.
description No Tells the parent agent when to delegate. Most important field.
model No fast, inherit, or a specific model ID.
readonly No true restricts write permissions.
is_background No true runs without blocking.

Agents vs commands

Use commands when... Use agents when...
Single-purpose, one-shot task Multi-step work needing context isolation
You want to invoke it explicitly You want automatic delegation
Quick, repeatable action Long-running or parallel work
Output goes directly to you Output feeds back into the parent agent

Creating a new agent

Add a markdown file to .cursor/agents/ with YAML frontmatter:

---
name: agent-name
description: When to use this agent. Be specific.
model: inherit
---

Your system prompt here. Keep it concise — long prompts make agents slower, not smarter.

Best practices:

  • Start with 2-3 focused agents, not dozens of generic ones
  • Invest in the description — it determines when delegation happens
  • Keep prompts short and specific
  • Avoid generic agents like "helps with coding"