Agency-agents in detail: Putting an entire team of AI experts into development tools

msitarzewski/agency-agents is an Agent role library for AI programming tools and multi-agent workflows. It is not a single application, nor a collection of ordinary prompts, but a set of AI expert definition files split by professional teams: engineering, design, marketing, sales, product, project management, testing, security, GIS, game development, finance, academic research and other roles have been made into reusable Agents.

Each Agent document typically contains identity, working methods, deliverables, success criteria and communication style. You can install it into tools such as Claude Code, Cursor, Codex, Gemini CLI, OpenCode, Aider, Windsurf, Kimi Code, Osaurus, Hermes, etc. You can also use these Markdown files as reference templates to transform them into your own team prompt words.

Who is the project suitable for?

This project is suitable for three types of people:

• I often use Claude Code, Codex, Cursor and other tools for development, and hope to switch professional roles quickly.
• If you want to build a multi-agent workflow, you need ready-made role divisions, boundaries, and delivery standards.
• I want to learn how to write high-quality Agent definition documents, including personality, process, constraints and output format.

It’s more like a “library of AI experts” than “magic software that does everything automatically.” The key to really using it well is to choose the right role, give it clear context, and incorporate the Agent’s output into your code review, testing, and release processes.

Installation method

It is recommended to use the desktop application to install the project. The officially provided Agency Agents App supports macOS, Linux and Windows. You can browse the complete Agent list and install it into Claude Code, Cursor, Codex, Gemini, Osaurus and other tools with one click, and is also responsible for updates.

macOS users can use Homebrew to install:

1

brew install --cask msitarzewski/agency-agents/agency-agents

If you prefer the command line, you can clone the repository:

1
2

git clone https://github.com/msitarzewski/agency-agents.git
cd agency-agents

Install to Claude Code:

1

./scripts/install.sh --tool claude-code

You can also copy only a certain category, for example, just the engineering directory:

1

cp engineering/*.md ~/.claude/agents/

Before installing to other tools, first generate the integration files required by each tool:

1

./scripts/convert.sh

Then install interactively:

1

./scripts/install.sh

You can also specify tools:

1
2
3
4
5
6
7
8
9

./scripts/install.sh --tool cursor
./scripts/install.sh --tool codex
./scripts/install.sh --tool gemini-cli
./scripts/install.sh --tool opencode
./scripts/install.sh --tool aider
./scripts/install.sh --tool windsurf
./scripts/install.sh --tool kimi
./scripts/install.sh --tool osaurus
./scripts/install.sh --tool hermes

If you don’t want to install all Agents at once, you can select precisely by division or agent:

1
2
3
4

./scripts/install.sh --tool claude-code --division engineering,security
./scripts/install.sh --tool cursor --agent frontend-developer,ui-designer
./scripts/install.sh --list teams
./scripts/install.sh --tool opencode --division engineering --dry-run

The README specifically reminds that OpenCode will only register about 119 Agents when it is currently running, and may be silently discarded after exceeding the limit. Therefore it is best to use --division to select a subset when installing OpenCode.

Usage

After installation, you can directly let the model switch roles in the AI ​​programming tool. For example:

1

Activate Frontend Developer mode and help me build a React component.

A more practical way to write it is to give the task, context and expected deliverables together:

1
2
3

Activate Backend Architect.
Review this API design, identify scaling risks, and propose a database schema.
Keep the output as decisions, trade-offs, and next implementation steps.

If you are working on a complex project, you can combine multiple Agents:

• Product Manager First organize the PRD and acceptance criteria.
• Software Architect Evaluates system boundaries.
• Backend Architect Design API and data structures.
• Frontend Developer implements the interface.
• API Tester and Reality Checker are used for verification.
• Technical Writer Supplementary documentation.

This type of combination makes it easier to control quality than “having a general assistant do it from start to finish” because the perspective from which each role is responsible is clearer.

Complete function introduction form

The following is a complete function list organized according to the roster structure of the README. The Agent name remains in English, and the capabilities and applicable scenarios are described in Chinese to facilitate quick screening.

real usage combination

The README gives several typical scenarios, and the idea is very straightforward: don’t assign all tasks to one role, but form teams according to project stages.

When making a Startup MVP, you can combine Frontend Developer, Backend Architect, Growth Hacker, Rapid Prototyper, and Reality Checker. The first two are responsible for implementation and architecture, the growth role is responsible for the customer acquisition path, and the testing role is responsible for pre-release quality.

When doing marketing activities, you can combine Content Creator, Twitter Engager, Instagram Curator, Reddit Community Builder, Analytics Reporter. This way each platform has a corresponding tone and execution focus.

When delivering enterprise functions, you can combine Senior Project Manager, Senior Developer, UI Designer, Experiment Tracker, Evidence Collector, and Reality Checker. This set is suitable for teams that need a closed loop of scope management, implementation, design, experimentation and acceptance.

When doing security or infrastructure work, it is not recommended to have non-security roles “drop in”. A better approach is to let Security Architect do the threat model first, then let Application Security Engineer or Penetration Tester do the code and attack surface inspection, and finally use Incident Responder to supplement the emergency process.

Usage suggestions

First, install a small amount first. There are many characters in this warehouse, and it is easy to find them difficult to find among the tools when they are full at once. For daily development, you can install Engineering, Testing, and Security first; for content and growth, you can install Marketing, Paid Media, and Sales.

Second, give the Agent clear boundaries. For example, if you ask Minimal Change Engineer to fix bugs, tell it not to refactor; if you ask Software Architect to evaluate the architecture, don’t ask it to directly change the code at the same time.

Third, treat Agent output as professional advice and don’t skip verification. Code needs to be tested, security conclusions need to be reviewed, and marketing and compliance recommendations need to be combined with actual business and local regulations.

Fourth, these Agents can be learned as writing templates. A good Agent is not just “you are an expert in such and such”, but should also include task boundaries, processes, delivery formats, quality standards and how to handle failures. A large part of the value of agency-agents lies in these structural settings.