Agency Agents

An AI Company in a Box -- Explained Like You're 5

agent37.club

What Is It?

🏢

Think of it like...

A pretend company full of AI employees, each with their own job title, personality, and to-do list.

Agency Agents is an open-source collection of 120+ specialized AI agent personas organized into departments -- just like a real company.

Created by msitarzewski and the community. 57.8k stars on GitHub.

Imagine a Whole Company...

...but every employee is an AI agent

👩‍💻

Engineering

23 agents: frontend, backend, DevOps, security, AI, mobile...

🎨

Design

8 agents: UI, UX, brand, visual storytelling...

📣

Marketing

26 agents: SEO, social media, content, growth...

💰

Sales

8 agents: outbound, deals, proposals, coaching...

🧪

Testing

8 agents: QA, performance, accessibility, API...

📦

+ 6 More Depts

Product, Project Mgmt, Support, Spatial, Specialized, Academic

What Makes an "Agent"?

Each agent is a detailed personality file with four parts

Identity -- A name, emoji, color, and vibe statement. Like a character sheet in a game.
Mission -- Their specific job responsibilities. A frontend dev builds web UIs. A growth hacker finds new users.
Workflow -- Step-by-step instructions for how they do their work. Like a recipe.
Deliverables -- What they actually produce. Real code templates, reports, strategies.

Key Insight

These aren't vague prompts. Each agent has real code examples, success metrics, and rules they must follow.

# Example: Frontend Developer
Name: Frontend Developer
Emoji: 🖥️
Vibe: "Pixel-perfect precision"
Rules:
  - Lighthouse > 90
  - WCAG 2.1 AA always
  - LCP < 2.5s

The Boss: Agents Orchestrator

One agent to rule them all

🎛️

Like a project manager who...

Reads the plan, assigns tasks to the right agents, checks their work, and loops back if it's not good enough.

Spec

→

Architect

→

Build

→

Test

→

Ship!

If a test fails? The orchestrator sends it back to the builder. Up to 3 retries.

How Do You Actually Use It?

Three ways, from easy to advanced

1. Copy-Paste

Browse the agent files on GitHub. Copy the parts you like into your own AI tool.

2. Claude Code Integration

Run the install script. Agents appear in your Claude Code session as ready-to-use personas.

3. Multi-Tool Install

Works with Cursor, Windsurf, Aider, Gemini CLI, and more.

# Install for Claude Code
./scripts/install.sh

# Or target a specific tool
./scripts/install.sh --tool cursor

# Convert formats first
./scripts/convert.sh

Why Not Just One Big Prompt?

One Giant Prompt

Tries to be everything at once
Loses focus on long tasks
No quality checks built in
Same "voice" for every task

Specialized Agents

Each expert focuses on one domain
Real workflows, not just vibes
QA loops catch mistakes
Different personalities for different jobs

"Specialization beats generalization"

Real-World Example

You say: "Build me a landing page for my startup"

🎛️

Orchestrator reads the spec, creates task list

🎨

UX Architect designs the layout and flow

🖥️

Frontend Developer builds it (responsive, fast, accessible)

✍️

Content Creator writes the copy

🧪

QA checks everything -- fails? Goes back to step 3

🚀

Ship it!

By the Numbers

120+

Specialized agent personas

11

Departments (divisions)

57.8k

GitHub stars

Open Source (MIT)

Free to use, modify, and share. Community-driven development with contributions from many authors.

Multi-Platform

Works with Claude Code, Cursor, Windsurf, Aider, Gemini CLI, and OpenCode.

The Big Idea

"Instead of one AI that does everything okay, have many AIs that each do one thing great."

github.com/msitarzewski/agency-agents

agent37.club