Someone Open-Sourced an Entire AI Agency

A free GitHub repo holds more than 230 specialist AI agents sorted into departments: engineering, marketing, sales, support, finance. Drop them into Claude Code or Cursor and stop making one chatbot do every job in your company. Setup takes about ten minutes and costs nothing.

You have one AI assistant. It writes your quotes, answers your support mail, drafts your marketing, and checks your spreadsheets. Same brain, four hats, and it loses the thread every time it switches.

No company runs like that. Companies hire specialists and give each one a defined role, a personality, and a way of working.

Someone built that structure for AI and put it on GitHub for free. It is a collection of specialist agent files, sorted by department like a real org chart. You install the ones you need, name the agent you want, and hand over a task. That sentence is the whole method. The rest of this article is the walkthrough: what the repo is, how to install it, which agents earn a seat in a business, and where it falls short.

What Is a Subagent?

A subagent is a text file that gives your AI assistant one defined job: a role, a voice, a workflow, and the deliverables it owes you. When you call it, the assistant stops being a generalist and works inside that one lane. One file, one specialist. Claude Code reads these files from a folder called ~/.claude/agents/. Cursor reads them as rule files in .cursor/rules/.

The Repo: An Org Chart in Markdown

The repo is agency-agents by msitarzewski, published as "The Agency". MIT license. The README's own words: "Use freely, commercially or personally. Attribution appreciated but not required."

When the reel about this repo went around Instagram in May 2026, it held 147 agents. The README now says 230+. Someone is maintaining this, and the roster keeps growing.

The agents sit in division folders. Here are the ones a business owner will open first, with the rough counts the README gives:

There are more divisions: game development, GIS, healthcare, academic research, spatial computing. If those words never appear in your P&L, skip those folders.

Open any agent file and you find the same skeleton: who the agent is, how it works, what it hands back, and what "good" looks like for that role. That structure is the difference between this and a folder of prompts. A prompt asks for an output. An agent file defines a job.

Setup: Three Steps to Your First Agent

Browse before you install. Open the repo on GitHub, click into sales/ or marketing/, and read two or three files top to bottom. You are about to hand these instructions to your AI. Read them the way you would read a CV.

Then pick your path.

If You Use Claude Code

Clone the repo and run the installer:

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

Or skip the script and copy only what you want:

cp sales/*.md ~/.claude/agents/

Then open Claude Code and name the agent: "Use the outbound strategist agent for this." Done. The specialist is on the clock.

If You Use Cursor

Same installer, different flag:

./scripts/install.sh --tool cursor

Each agent lands as a .mdc rule file in .cursor/rules/, and Cursor picks it up from there.

If You Use Something Else

The install script converts the agents for Codex, GitHub Copilot, Gemini CLI, Aider, Windsurf, OpenCode, and several more. Run ./scripts/install.sh --interactive and it walks you through it. The agents are plain text. They run wherever your model runs, so you are not locked to any one vendor.

Copy This: The Roster Prompt

Do not guess which agents you need. Make your AI pick them against your real week.

I run a [BUSINESS TYPE: residential electrical company] with
[TEAM SIZE: 9] people. Most of my week goes to
[TOP 3 TASKS: quoting jobs, chasing unpaid invoices, writing
follow-up emails].

I am looking at the agency-agents repo on GitHub. It has divisions
for engineering, design, marketing, sales, product, testing,
security, support, and finance.

Pick the five agents I should install first. For each one tell me:
1. The task in my week it takes over
2. What I hand it on day one
3. What I should get back

Do not pick agents for work I do not do. Five picks maximum.

Fill the brackets, paste it into Claude or whatever you run, and install only what comes back. Five files, not 230.

Your First Real Run

Installed? Give one agent one real task from this week. Not a test question. A task.

Use the [AGENT NAME: outbound strategist] agent.

Here is a real task from this week: [TASK: write the follow-up
for the kitchen remodel quote I sent Tuesday. The client went
quiet.]

Context: [PASTE THE EMAIL THREAD OR YOUR NOTES]

Give me the finished deliverable, not a plan for making it.

If the output is wrong, say exactly what is wrong: "the second paragraph sounds like a template, rewrite it in the same tone as my pasted emails." Specialists take direction. That is what they are for.

Where This Fits in a Business

• Sales follow-ups. The sales division has agents for outbound, discovery, and proposals. Feed one your pipeline notes and get the drafts you keep postponing.
• Marketing that ships. Thirty agents cover content, SEO, and email. One agent per channel beats one assistant juggling all of them.
• Finance sanity checks. The bookkeeping and analysis agents read your numbers before your accountant bills you to do the same.
• Support replies. The support agents draft responses in a consistent voice instead of whatever tone you had at 11pm.
• A second set of eyes. The testing division includes a reality checker and an evidence collector. Point one at your own claims before a client does.

The Mistake That Makes It Fail

Installing everything. The repo has 230+ agents and the installer will happily give you all of them. Now your AI has a phone book instead of a team, you have no idea who does what, and every session drags a crowd of irrelevant specialists behind it.

Nobody staffs a nine-person company with 230 hires. Pick five. Fire the ones that do not earn their seat. Add one when a real task demands it. Treat the repo as a talent market, not a checklist.

Honest: The Limits You Should Know

1. These are files, not employees. An agent knows its craft, not your business. Until you feed it your context, your pricing, your tone, and your clients, it gives you competent generic work.
2. No hands. The sales agent cannot send an email and the finance agent cannot open your bank account. Connecting agents to real tools is separate work, and it is where the real leverage lives.
3. The roster tilts toward software. Engineering, game development, and GIS make up a large share of the 230+. A non-technical business will use maybe a tenth of the collection. That tenth is still free.
4. Quality varies by file. This is a community-maintained collection under MIT, not a certified product. Read every file before you install it, because it is literally instructions your AI will follow.
5. The model still does the work. An agent file shapes behavior. It does not upgrade the underlying model. Weak model, weak specialist.

Do This Today

If you only do one thing this week: run the roster prompt, install the five agents it picks, and give one of them a real task before you close the laptop today.

If the output is useless, you spent ten minutes and zero dollars. If it is decent, you just found out what an org chart of specialists feels like, and you will not go back to one overworked chatbot.