The Agency: A Company That Lives Inside Your Hard Drive — 232 AI Employees, 16 Departments, Zero Sleep

Forum topic · ✨步子哥 · 2026-07-01

Summary

The Agency is an open-source GitHub repository by American developer Michael Sitarzewski containing 232 AI agent "role cards" organized into 16 departments — engineering, design, testing, security, marketing, GIS, game development, and more. Each agent is defined as a complete persona in a Markdown file with identity, mission, rules, workflow, and communication style, rather than a simple prompt. The project compiles these "master" files into formats for 14 AI coding tools including Claude Code, Cursor, GitHub Copilot, and Kimi Code via a convert.sh script acting like a master-to-disc duplication pipeline. Strict automation enforces single-source-of-truth consistency across divisions.json, directory structure, and scripts. Notably, roughly half of the 36-person marketing department targets Chinese platforms (Xiaohongshu, Zhihu, Douyin, Bilibili), largely contributed by community member hedonnn — reflecting a philosophy that the Chinese internet is a parallel ecosystem, not a translation of the Western web. The project is entirely free under MIT license, funded by GitHub Sponsors, with a desktop app at agencyagents.app. Launched in October 2025 from a Reddit discussion with 51 agents and 9 departments, it grew to 232 agents across 357 commits in 8 months. Its core bet: "persona"-style AI specialization will outlast module-based "skill" standards like SKILL.md.

A Company That Lives Inside Your Hard Drive: The Agency

Overview

The Agency is an open-source GitHub repository by Michael Sitarzewski (VP of Innovation & Technology at Tansan-era firm Tandem Theory) that functions as a company living entirely in a Git repository: 232 AI "employees" across 16 departments, each defined as a detailed role card in Markdown.

Key points

Structure: 232 agent persona files organized into 16 departments (Engineering: 31, Marketing: 36, Specialized: 40+, GIS: 13, Sales: 10, Security: 10, Design: 9, Testing: 9, PM: 7, Paid Media: 7, Finance: 5, Academic: 5, plus game dev, spatial computing, and more).
Role cards, not resumes: Each agent is a complete "working persona" with identity, core mission, key rules, workflow, and communication style — e.g., a front-end engineer card instructs the AI to ask about use cases and accessibility before writing React code, and to flag known Safari pitfalls. Roles include quirky ones like the Whimsy Injector (design) and Reality Checker / Evidence Collector (QA).
Persona vs. Skills: The project bets on "personification" against the mainstream "skill-ification" trend (Anthropic's SKILL.md standard, adopted by 40+ tools). The philosophy: a recipe can be copied, but a chef's feel cannot. Skills = recipes; personas = chefs.

Engineering discipline

divisions.json as constitution: Every commit must keep four places perfectly aligned — disk directories, the conversion script's department array, the check script's array, and CI path config. A mismatch blocks the merge ("Single Source of Truth").
Three security gates: check-tools (validates configs for 14 supported tool platforms), check-divisions (department alignment), and lint-agents (agent file quality). The first two are pure Bash with zero dependencies, so they run identically everywhere.
Master-and-burn architecture: convert.sh compiles the 232 Markdown "master" files into formats for 14 tools (Claude Code, Cursor, Codex, Gemini CLI, GitHub Copilot, Aider, Windsurf, Kimi Code, Qwen, etc.) via 10 shared renderers. All output in integrations/ is stamped "generated — do not edit manually."
Clever workarounds: OpenCode has a bug that silently drops agents beyond ~119; The Agency can't fix OpenCode but warns users at install time and suggests the --division flag. A .gitattributes forcing LF line endings prevents invisible \r characters from breaking frontmatter parsing.

The Chinese squad

Roughly half of the 36-person marketing department targets Chinese platforms natively: Xiaohongshu ("seeding" content), Zhihu, Bilibili, Douyin, Kuaishou, Weibo, WeChat public accounts, private-domain traffic, livestream commerce. These aren't translations — the Chinese internet is a "parallel universe" where concepts like 私域 (private domain) and 素人 (amateur creator) seeding have no Western equivalent. Much of this was contributed by Chinese community member hedonnn, with a dedicated CONTRIBUTING_zh-CN.md guide — "community as localization."

History and business model

Origin: October 13, 2025, sparked by a Reddit debate on whether AI agents should specialize rather than be generalists. First version: 51 agents, 9 departments. Eight months later: 232 agents, 16 departments, 357 commits, multiple contributors, and a desktop app (agencyagents.app).
Free forever: MIT-licensed agents AND a fully open-source desktop app (no telemetry, no accounts, no paywall), funded by voluntary GitHub Sponsors. The author holds a day job, so there's no pressure to monetize. The real bet is becoming the canonical standard for AI agent persona definitions — like Linux for servers.

What it means

Human-AI collaboration is shifting from one assistant per person to managing a team of specialists. Rather than prompting, you deploy complete working personas. And "how to organize many specialists efficiently" — a problem humanity solved over centuries with departments, roles, and QA — is now being ported wholesale into the digital world. One person, one Git repo, and strangers on the internet can now run a 232-employee agency. In the AI era, "a one-person company" is no longer just a metaphor.