msitarzewski/agency-agents — Agency Agents: A Distribution System for AI Agent Personalities
Transcript
QuickFacts
Welcome to Agency Agents — a project that's all about giving AI assistants personality and expertise. This is a catalog and distribution system for specialized agent prompts that work across sixteen different coding tools.
PlainEnglish
At its heart, this is a staffing agency for AI. You've got two hundred sixty-three specialized agent personalities — each one an expert in something narrow like frontend development, Reddit marketing, or reality-checking QA.
PlainEnglish
These agents are organized like specialized staffing firms. Engineering has fifty-four agents covering frontend, backend, DevOps, AI, mobile. Marketing has thirty-six for SEO, social media, content — even China platforms. Specialized has fifty-six covering everything from customer service to workflow architecture. And there are fourteen more divisions for design, sales, security, GIS, game development, and more.
Architecture
Each agent is a Markdown file with a consistent structure. Up top, you've got YAML frontmatter with the basics — name, description, color, emoji. Then the body is organized into sections: who they are, what they do, rules they follow, examples of their output, their workflow process, and how to measure success. It's like a living resume for an AI personality.
PlainEnglish
Here's the thing — not every tool speaks the same language. The agents are written in Markdown with YAML frontmatter, which Claude Code reads natively. But Cursor wants mdc rule files, OpenClaw wants a three-file workspace split by personality and operations, and Aider wants everything concatenated into one conventions file. So the repo includes a conversion pipeline that acts like a translation service, turning one source format into sixteen different dialects.
Architecture
The conversion pipeline is a bash script called convert dot sh. It reads all two hundred sixty-three agent files from the division directories, then runs them through sixteen different converter functions — one for each tool. Some converters write one file per agent, others accumulate everything into a single file, and one even builds a Python plugin artifact. The outputs land in an integrations directory, organized by tool.
Architecture
Once you've got the integration outputs, the installation system takes over. It's like a package manager for agents. First, it detects which tools you have installed by checking for directories like dot claude or running version commands. Then it shows you an interactive checkbox UI to pick exactly what you want. If the integration outputs are missing or stale, it auto-runs the conversion pipeline. Finally, it copies the generated files into the right places — dot claude slash agents, dot cursor slash rules, and so on. Now your tools see the agents and you can activate them by name.
Architecture
Two JSON files serve as configuration contracts. divisions dot json defines the seventeen divisions with labels, icons, and colors. tools dot json defines the sixteen integration targets with detection rules, destination paths, and format expectations. The installer reads divisions dot json to populate its team selection screen. The converter reads tools dot json to know what formats to generate. And CI validation scripts act like a type checker — if you add a division but forget to update the JSON, the build fails.
PlainEnglish
Let's walk through a real scenario. A developer wants agents in Cursor, so they run install dot sh with the tool flag set to cursor. The installer checks if the integration outputs exist — if they're missing or stale, it calls the converter to generate fresh mdc files. Then it copies those files into dot cursor slash rules in the current project. Next time the developer opens Cursor, all the agents are sitting there ready to be activated.
Community
This is a very young project — the last commit was just three days ago, and there's one contributor so far, Michael Sitarzewski. It's actively being built, with recent commits and two branches in flight. No tags yet, which tells us it's still in early development. This is a solo effort at the moment, but the infrastructure is solid and ready for contributions.
PlainEnglish
So that's Agency Agents. It's a toolkit for giving AI assistants real expertise. You've got two hundred sixty-three specialist personalities across seventeen divisions, a conversion pipeline that speaks sixteen tool dialects, a bash-based installation system with an interactive wizard, metadata contracts enforced by CI, and even a companion desktop app for one-click installs. It's modular, extensible, and built for developers who want their AI to be more than a general-purpose assistant. Thanks for watching.
How this was made
Lenzon read msitarzewski/agency-agents and generated this walkthrough automatically. The narration above is the transcript of what it says.
Explain a pull request from your own repo
Point Lenzon at a repo or a pull request and get a narrated walkthrough like this one.
Try it