Agency Agents

A subagent is a scoped role you hand a piece of work to: its own instructions, its own tools, its own slice of context. Writing a good one takes a couple of hours of iterating on the prompt. Agency Agents is a library of 230+ of them already written, organised into seventeen divisions — Engineering, Design, Testing, Security, Product, Marketing, Finance, Game Development, GIS, Healthcare, and more.

Each agent file is more than a persona line: front matter with identity and mission, critical rules, a worked process, concrete deliverables with examples, and success metrics. That structure is the actual value — it’s the part people skip when they write their own.

Install link

The repository ships an installer that detects which tools you have and lets you choose by division or by individual agent:

  git clone https://github.com/msitarzewski/agency-agents.git
cd agency-agents
./scripts/install.sh --tool claude-code
  

Run it with no flags for the interactive picker. For editors other than Claude Code, convert first:

  ./scripts/convert.sh
./scripts/install.sh
  

Claude Code, Copilot, Cursor, Windsurf, Aider, Gemini CLI, OpenCode, Codex and others are supported. There’s also a desktop app if you’d rather not touch the terminal. MIT licensed, so adapting an agent for your team is fine.

Claude Code reads agents from ~/.claude/agents/ (yours) and .claude/agents/ (the project’s) — the installer writes to the first; copy the handful your team shares into the second and commit them.

Use one link

  > Use the Frontend Developer agent to review this component for accessibility
> Have the Security Auditor agent look at the auth middleware
  

The agent runs with its own context, so a long review doesn’t crowd out your main session — which is the practical reason to bother, quite apart from the expertise in the prompt.

Install six, not two hundred link

This is the part the star count won’t tell you. A wall of 230 agents makes every selection worse: you can’t remember which exist, the picker becomes noise, and an agent you never invoke is a file that will drift out of date. OpenCode makes the point concretely — its runtime registers roughly 119 agents and silently drops the rest, so targeting it means --division flags whether you like it or not.

A workable approach:

1. Install one division you actually work in — Engineering, say.
2. Use it for a week and notice which three or four agents you invoke by name.
3. Uninstall the rest, and copy the survivors into the project’s .claude/agents/ so your team gets them too.
4. Edit them. These are starting points; the version that knows your stack, your test command, and your conventions is worth five generic ones.

Read before you run link

The agent definitions are prompts from a third party, and a prompt is instructions your agent will follow. Skim the ones you install — particularly any that tell the agent to run commands, install packages, or reach the network. That’s not a knock on this project; it’s the correct habit for every prompt library you install.

Next link

The most unexpected thing on this list — an agent that edits video → OpenMontage