What it is
Agency Agents is a collection of specialized AI roles described in Markdown. Each role defines how an assistant should work: mission, communication style, expected deliverables, process, and success criteria.
The project targets tools such as Claude Code, Cursor, Codex, Aider, Windsurf, Gemini CLI, and others. The point is not a magical model; it is giving the assistant a sharper role than simply “help with code.”
How it appeared and why it stuck
The project grew from the practical problem that one broad assistant quickly becomes too vague. Code review needs one mode of thinking, design another, security another, documentation another. Agency Agents splits those modes into separate files.
Its popularity reflects how many developers started building their own instruction sets for AI tools. Here that setup is public and comes with installation scripts.
What is inside
The repository contains groups of agents, install and conversion scripts, variants for different tools, and manual usage notes. Users can install all roles or pick specific divisions.
Installing selected roles
This example shows how to attach the set to a specific tool and choose only the needed directions, not how the agents themselves work.
./scripts/install.sh --tool claude-code --division engineering,security
./scripts/install.sh --tool cursor --agent frontend-developer,ui-designer
./scripts/convert.shWhere it helps
Agency Agents helps people who already work with AI assistants and want to explain the role less often. Instead of a long manual request, they can select a prepared profile: interface developer, code reviewer, security specialist, documentation writer.
For a team, this approach can create shared language: members understand which assistant mode is being used and what output is expected.
Strengths and limits
The strength is portability. Markdown files are easy to read, edit, version, and move between projects.
The limitation is that output quality still depends on the model, context, and human review. A role guides the assistant; it does not automatically make it an expert.