The Agency: An Open-Source Company With 232 AI Employees Living in Your Hard Drive

Forum topic · ✨步子哥 · 2026-07-01

Summary

The Agency is an MIT-licensed GitHub repository by American developer Michael Sitarzewski that functions as a "company stored on a hard drive": 232 AI agent persona files (role cards) organized into 16 departments, from engineering and security to GIS, game development, and a 36-person marketing division. Roughly 20 of those marketing agents natively cover the Chinese internet ecosystem—Xiaohongshu, Zhihu, Bilibili, Douyin, Kuaishou, WeChat, private-domain traffic—built by Chinese community contributors rather than translated from Western concepts. The project's engineering discipline includes a divisions.json acting as a single source of truth enforced by three automated CI checks written in dependency-free Bash, a "master-to-burn" converter (convert.sh) that compiles the same Markdown role cards into formats for 14 AI coding tools (Claude Code, Cursor, Codex, Gemini CLI, Copilot, and more), and strict LF line-ending rules to prevent parser breakage. Started in October 2025 from a Reddit discussion with 51 agents in 9 departments, it grew to 232 agents in about 8 months via community contributions. It is entirely free—no subscriptions, no paid tiers—sustained by GitHub Sponsors and a day job, betting that persistent "personas" of AI experts will outlast a skills-module approach like SKILL.md.

A Company That Lives in a Git Repository

At 3 a.m., you need three people: a front-end engineer to debug a React component misaligning in Safari, an SEO expert to audit your page structure, and a copywriter to polish a slogan. Instead of hiring, you open a folder containing 232 neatly filed "expert profiles," pick three, drop them into your AI coding tool, and they start working. No interviews, no salaries, no schedules.

This folder is The Agency, an open-source, free GitHub repository by American developer Michael Sitarzewski. Its 232 "employees" are organized into 16 departments like a real agency—with one difference: the entire company is a Git repository.

The Roster: 232 Agents, 16 Departments

Engineering (31): front-end, back-end, mobile, AI engineers, DevOps, plus unusual roles like a "Codebase Onboarding Engineer," a "Minimal Change Engineer" (change one word, never two), and an "Incident Response Commander."
Design (9): includes the Whimsy Injector, whose job is adding personality and delight to cold interfaces.
QA (9): a Reality Checker acting as quality gate and an Evidence Collector for screenshot-based proof.
Security (10), Academic (5: anthropologist, geographer, historian, narratologist, psychologist for world-building), Game Development (Unity, Unreal, Godot, Roblox, Blender), Spatial Computing (Vision Pro, WebXR), GIS (13), plus finance, project management, paid media, and sales.
Marketing (36): half for global platforms (Twitter, Instagram, TikTok, LinkedIn, Reddit), half natively covering the Chinese internet—Xiaohongshu, Zhihu, Bilibili, Douyin, Kuaishou, Weibo, Baidu SEO, WeChat, live commerce, private-domain operations.
Specialized (40+): MCP protocol builders, CFO, HR, real estate, M&A, organizational psychology, ESG, data privacy.

Each agent has an "ID badge": a name, an emoji icon, and a vibe field describing its temperament.

Role Cards, Not Resumes

A generic AI assistant lacks a "stance." The Agency's agents are role cards: engineered definitions containing identity, core mission, key rules, workflow, and communication style. A front-end role card doesn't just "know React"—it asks about context and accessibility, then delivers a complete plan with known Safari caveats.

The project's philosophical bet is personalization over skills-ification. While Anthropic's SKILL.md open standard (adopted by 40+ tools) slices AI capability into reusable skill modules, The Agency's unit is a complete expert. The analogy: a skill is a recipe; a persona is a chef. Recipes copy easily; a chef's intuition doesn't. As models internalize skills, "a colleague with a stable working personality" may be the more durable abstraction.

divisions.json: The Company Constitution

With 232 agents, organization requires discipline. divisions.json defines all 16 departments, and an automated check verifies four places stay in perfect sync on every commit:

1. Actual directories on disk 2. The department array in the conversion script 3. The department array in the check script 4. Path configuration in the CI workflow

Any mismatch blocks the merge—Single Source of Truth, enforced. A directory counts as a department only if it contains actual agent Markdown files (the strategy folder, full of playbooks, is excluded). A folder name alone doesn't make a department; real employees do.

Three "security gates" run in CI:

check-tools: validates configs for all 14 supported tool platforms
check-divisions: the four-way department alignment check
lint-agents: validates every changed agent file's name, description, and format

The two always-on gates are written in pure Bash with zero dependencies—not even jq—so they run identically on macOS terminals and cloud servers.

Masters and Burns: One Card, 14 Tools

The Agency supports 14 tools: Claude Code, Cursor, Codex, Gemini CLI, GitHub Copilot, OpenCode, OpenClaw, Antigravity, Aider, Windsurf, Kimi Code, Osaurus, Hermes, and Qwen. Each demands different agent file formats—some Markdown, some JSON, some one file per agent, some one giant roster.

Hand-writing 232 × 14 files is unsustainable, so convert.sh acts as a compiler. The 16 department Markdown files are the master tapes—the only hand-written originals. Everything in integrations/ is a "burned copy," stamped "generated by convert.sh, do not edit." Change a master, re-burn.

Key mechanisms:

A format field defines a renderer contract: only tools with byte-identical output share a renderer. 14 tools share 10 renderers.
An installKind field defines three install mechanisms: per-agent (one file per agent), roster (all 232 agents merged into one file, as Aider's CONVENTIONS.md or Windsurf's .windsurfrules), and plugin (a single installable package for Hermes).
For OpenClaw, which requires three files per agent (SOUL.md, AGENTS.md, IDENTITY.md), a clever renderer splits each master by section-heading keywords.

Known limits are handled honestly: OpenCode's runtime silently drops agents beyond ~119, so the installer warns you and suggests --division to install a subset. And a one-line .gitattributes (*.md text eol=lf) forces LF line endings everywhere, because a stray Windows CRLF \r turns --- into ---\r and breaks frontmatter parsing with baffling errors.

The Chinese Squad: Native, Not Translated

Why does an American developer's repo contain ~20 agents fluent in "种草" (grass-planting), "老铁," and private-domain traffic? Because the Chinese internet is a parallel universe, not a translation:

Private-domain traffic (私域流量): no Western equivalent—it means keeping users in your own pond (WeChat groups, mini-programs).
Xiaohongshu seeding: not influencer marketing; it's built on ordinary users' authentic notes, long-tail search positioning, and content matrices—not KOL buys.
Douyin's algorithm differs from TikTok's despite shared parentage.

A universal "social media strategist" agent fails in China not on language but on underlying mechanics—like an international chess master sitting down at Chinese chess. Notably, these agents were largely contributed by Chinese community members (e.g., hedonnn), passing originality checks, with a maintained Chinese contribution guide (CONTRIBUTING_zh-CN.md). It's "community-as-localization": the author provides framework and quality control; locals bring native knowledge.

From a Reddit Post to a 232-Person Team

The first commit appeared on GitHub on October 13, 2025: 51 agents, 9 departments. Inspired (per Chinese tech community reports) by a Reddit debate on whether AI agents should specialize rather than generalize, Sitarzewski—a serial founder of 8 companies, Techstars Cloud alumnus, Dallas Startup Week founder, and VP of Innovation & Technology at Tandem Theory by day—built the first version. In about 8 months: 357 commits, 232 agents, 16 departments, multiple contributors, and even a desktop app (agencyagents.app) with graphical browsing, one-click install, and drift detection for non-engineers. Some commits credit collaboration with Claude—a person using AI to build an AI company for humans.

How It Stays Free

The structure is three layers: the MIT-licensed agent catalog, an MIT-licensed desktop app (no telemetry, no accounts, no paywall), and a voluntary GitHub Sponsors link. This is sponsorware—the street musician model. Sitarzewski funds it with his day job; sponsors buy no features, just support. No pricing page, no login, no subscription exists anywhere. The deeper bet: become the canonical standard for AI agent persona definition—the way Linux and HTTP created value beyond direct monetization.

What It Means

Three shifts:

1. Human-AI collaboration is moving from one-to-one to one-to-many. You're no longer prompting an assistant; you're staffing and managing a team. 2. "Role" beats "prompt."" A complete working persona produces predictable results; a one-line instruction gambles on luck. 3. Human organizational wisdom is being ported to AI.** Departments, role definitions, quality gates, process discipline—centuries of lessons in organizing people now live in Markdown and JSON. A future "AI organizational theory" may differ, but until then, this is a pragmatic starting point.

Will the persona bet beat the skills bet? Short-term, likely yes—models are smart, but "smart" and "having a stable working personality" differ. Long-term, if models can generate stable personas on demand, preset cards may fade. For now, a tame mount beats a fast horse for most riders.

The simplest fact remains: one person, one Reddit post, 8 months, 232 agents, 16 departments, 14 platforms, a Chinese contribution guide, a desktop app. No funding, no company, no marketing budget. A one-person company in the AI age is no longer a metaphor. Its lights are always on—not from overtime, but because it never needs to sleep.