The Agency: A 218-Strong Library of Specialized AI Subagents

Most agent setups start the same way: you keep re-explaining the same role to the model. “Act like a senior backend architect.” “Review this like a security engineer.” “Write docs like a technical writer.” Each of those is a reusable persona you end up rebuilding from scratch in every project.

agency-agents — “The Agency” — is a large, open-source library that ships those personas pre-made: 218 specialized subagents across 15 divisions, each a markdown file you drop into Claude Code (or a dozen other tools) and activate on demand. This post covers what it actually is, how to install and use it, how it compares to rolling your own, and where it falls short.

What it is

agency-agents (by Matt Sitarzewski, MIT-licensed) is a curated collection of AI agent definitions — not a tool or a framework, but content. Per the project, it’s 218 agents across 15 divisions (Engineering, Design, Sales, Marketing, Product, Testing, Security, Finance, Game Development, Academic, and more), grown from a Reddit thread into a community-maintained roster.

Each agent is a single markdown file describing a specialist: its identity and “personality,” its core mission and workflows, the concrete deliverables it produces (with code examples), and success metrics. You install the ones you want and invoke them as subagents — e.g. “activate Frontend Developer mode and help me build a React component.”

Why it matters

Breadth, instantly. 218 ready-made specialists means you rarely have to write the persona yourself — from Backend Architect and Code Reviewer to PPC Strategist and Whimsy Injector.
Tool-portable. The same agents install across Claude Code, GitHub Copilot, Cursor, Aider, Windsurf, Gemini CLI / Antigravity, OpenCode, OpenClaw, Codex and Kimi Code via a converter script — so your roster follows you between tools.
Selective install. You don’t have to take all 218: install by division or by individual agent, so you only load what’s relevant to your work.
Specialist, not generic. Each file is a deliberately deep, opinionated role (workflows, metrics, code patterns), which tends to steer the model better than a one-line “act as a…”.

How it works

There’s no runtime here — an “agent” is just a structured prompt. The library leans on the host tool’s subagent mechanism (Claude Code’s ~/.claude/agents/, and the equivalent for each other tool). You install the markdown files into the right directory, then the model loads a given persona’s instructions when you activate it, and answers in that role for the rest of the task.

Because they’re plain markdown, you can also use them as reference: open the file, copy the workflow or the deliverable template, and adapt it — no installation required.

Getting started

For Claude Code, the install script copies the agents into your Claude config:

# install all agents for Claude Code
./scripts/install.sh --tool claude-code

# or just one division
./scripts/install.sh --tool claude-code --division engineering,security

# or hand-pick agents
./scripts/install.sh --tool cursor --agent frontend-developer,ui-designer

For other tools, generate the integration files first, then run the interactive installer (it auto-detects what you have):

./scripts/convert.sh        # build integration files for all supported tools
./scripts/install.sh        # interactive wizard: pick tools + teams
./scripts/install.sh --list teams   # see every team + agent count

Then activate an agent in a session by name (“activate the Backend Architect”).

In practice

A couple of realistic uses:

A code review gate. Install the Code Reviewer and Security agents, and route PR diffs through them for a structured, role-consistent review instead of an ad-hoc “look at this.”
A cross-functional sprint. Pull the Frontend Developer, Backend Architect, and UX Researcher into the same project so each task gets answered in the right voice — and reach for a Finance or Paid Media agent when the work leaves engineering.

The point isn’t novelty — you could prompt all of this by hand — it’s that the personas are written, versioned, and shareable, so the behavior is consistent across people and projects.

How it compares

agency-agents sits in the “agent personas / subagents” space, and the useful contrast is what you’d otherwise do:

Option	What you get	Tradeoff
agency-agents	218 ready personas, multi-tool, install a subset	Quality varies; you curate which to keep
Write your own subagents	Exactly fits your stack and conventions	You build and maintain every one
A single focused skill (e.g. Ponytail, karpathy-skills)	One sharp behavior, easy to reason about	Narrow — one job, not a roster
A link directory (e.g. awesome-claude-code)	Pointers to find things	You still assemble and install yourself

The honest framing: agency-agents trades depth-per-agent and curation for breadth and convenience. If you want one well-tested behavior, a focused skill is cleaner; if you want a whole bench of specialists without writing them, this is the fastest way to get one — then prune.

Tradeoffs

Quality and consistency vary. 218 community-contributed personas won’t all be equally deep or current. Treat the roster as a starting point you curate, not a guaranteed-good team.
More agents isn’t better. Installing everything can overwhelm a tool: the project itself warns that OpenCode currently registers only ~119 agents and silently drops the rest (an upstream bug), which is exactly why selective install exists. Even where it works, a giant roster is its own discovery problem.
Personas aren’t capabilities. These are prompts, not tools or tested code — they shape behavior, and the actual output quality still depends on the underlying model. A confident “personality” can read as substance when it isn’t.
Drift. Markdown personas can fall behind current best practices or your codebase’s conventions, and nothing pins them to your stack.

Takeaway

agency-agents is the fastest way to get a broad bench of specialized subagents without writing them yourself — 218 personas, MIT-licensed, installable across most agent tools, and easy to take in part. Use it when you want range and portability and you’re willing to curate; reach for a single focused skill when you want one sharp, well-understood behavior instead. Either way, install the few you’ll actually use, not all 218.