agency-agents is a “complete AI agency” you drop into your coding agent: 232 specialized roles, defined in Markdown, organized into 17 divisions from Engineering to Marketing to Security. It is not software you run. Each agent is a prompt definition with a personality, a process, and concrete deliverables, and you install the ones you want into Claude Code, Copilot, Cursor, Antigravity, and eight other hosts. Understanding that distinction (a curated role library, not an executable framework) is the key to using it well.
What is actually in it
The 232 agents sit in division folders. The largest are Marketing (around 36 roles, including channel specialists for TikTok, Xiaohongshu, and livestream commerce), Engineering (around 33, with specialists like Backend Architect and Solidity Smart Contract Engineer), and a catch-all Specialized division (around 53 cross-domain roles such as Chief of Staff and M&A Integration Manager). Newer divisions like GIS (added 2026-06) show the library is still growing. Each role file carries YAML front matter plus sections for identity, mission, technical deliverables, workflow, and success metrics, so an agent gets a fully-shaped persona rather than a one-line prompt.
Install
Installation is a script-driven copy into your host’s agent directory:
# Claude Code, all engineering agents
./scripts/install.sh --tool claude-code
# Interactive wizard, pick tool and divisions
./scripts/install.sh
# Selective, e.g. Cursor with two divisions
./scripts/install.sh --tool cursor --division engineering,securityA convert.sh step generates host-specific formats for all 12 supported tools (Claude Code and Copilot take the raw .md; Cursor gets .mdc rules; Aider compiles to a single CONVENTIONS.md; Codex takes TOML, and so on). After install, you activate a role in conversation, for example “Activate Frontend Developer and help me build a React component.”
The limit that bites
Here is the practical gotcha you will not find by skimming the README. The project documents that OpenCode silently registers only about 119 agents and drops the rest (an upstream limit). If you install all 232 into OpenCode, agents past the cap vanish without error. The fix is to filter with --division so you stay under the ceiling. This is the single most important operational note for anyone running it on OpenCode, and it is easy to miss.
When it fits, and when it does not
It fits if you want a broad, ready-made cast of specialist personas and you work across more than code (the Marketing and Sales divisions are unusually deep). It fits less well if you want a small, sharp set of engineering quality gates: 232 roles is a lot of surface area, and many teams will only ever use a dozen. Because the agents are prompt definitions, their value rises and falls with the host model and with how disciplined the personas are; there is no runtime enforcing them.
How it compares
| Project | Shape | Note |
|---|---|---|
| msitarzewski/agency-agents | 232 role personas, 17 divisions | Breadth across business + engineering |
| obra/superpowers | Skills framework + methodology | Process discipline over role breadth |
| anthropics/skills | Official agent skills | Curated, first-party |
| addyosmani/agent-skills | Production engineering skills | Quality gates, narrower scope |
The contrast is breadth versus discipline. agency-agents is the widest cast of characters; superpowers and addyosmani/agent-skills are tighter sets focused on engineering rigor; anthropics/skills is the first-party baseline.
On the growth and the maintenance
The repo crossed six figures of stars on what looks like organic growth, with the README noting roots in a Reddit thread rather than a single launch event. Development is active: hundreds of commits in 2026, with new divisions landing through June. There is one honesty note worth flagging: the Security division ships agents that include smart-contract exploit demonstrations (reentrancy, flash-swap patterns) as teaching material, and the project scrubbed some hard-coded test credentials in mid-2026. Treat the security content as educational, and follow the project’s own guidance not to store API keys inside agent files.
FAQ
Is agency-agents free? Yes. agency-agents is MIT-licensed, and the only cost is the model usage of whatever host agent you run it in.
How many agents does agency-agents have? 232 specialized roles across 17 divisions. Note that OpenCode registers only about 119 and silently drops the rest, so filter with --division to stay under that cap.
Does agency-agents work without Claude Code? It needs a host agent, but not specifically Claude Code: it installs into 12 tools including Copilot, Cursor, Antigravity, and OpenCode. It is not a standalone program.
Are agency-agents real agents or just prompts? They are richly-structured Markdown personas (identity, mission, deliverables, workflow, metrics), so think of them as carefully written role prompts, not an autonomous runtime that executes on its own.
Related reading
If you want fewer, sharper engineering skills instead of a 232-role cast, compare addyosmani/agent-skills and obra/superpowers. For the first-party baseline, see anthropics/skills.