Run Your Own AI‑Powered Company with 170+ Ready‑to‑Work Agents

Repository Overview

The GitHub repository https://github.com/msitarzewski/agency-agents provides a collection of 177 AI agents organized into 17 departments, simulating a full‑staff AI company.

Agent File Structure

Each agent is defined in a Markdown file with four sections: identity & personality, core mission & workflow, technical deliverables (code examples, templates, outputs), and success metrics & communication style. This turns a prompt into a role‑based “senior colleague.”

Department Highlights

Engineering (29 agents) : frontend, backend, mobile, AI engineer, DevOps, SRE, embedded firmware, Solidity smart‑contract, Filament PHP, WeChat Mini‑Program, Feishu integration, etc.

Design (8 agents) : includes a “Whimsy Injector” that adds easter eggs, micro‑interactions, and surprises to products.

Paid Media (7 agents) : PPC strategist, keyword analyst, account auditor, GTM/GA4 conversion tracker, Performance Max creative, programmatic buyer, cross‑platform social planner.

Sales (8 agents) : MEDDPICC qualification scorer, SPIN questioning designer, sales‑engineer pre‑sales, RFP writer, pipeline‑funnel analyst.

Marketing (30 agents) : roles for Chinese platforms (Xiaohongshu, Zhihu, Bilibili, Douyin, Kuaishou, WeChat Official Accounts) and an “AI Citation Strategist” that optimizes brand citation rates in ChatGPT, Claude, Gemini, Perplexity.

Testing (8 agents) : includes a “Reality Checker” that enforces evidence‑driven release gates.

Other departments: Specialized (41 agents), Academic (5), Finance (5), Game Development (5), Spatial Computing (6), Strategy (3), plus Product, Project Management, and Support.

Installation

The repository consists of pure Markdown files. For Claude Code (officially recommended), clone the repo and run the install script:

1、克隆项目
2、进入项目目录 cd agency-agents
# 一键全装
./scripts/install.sh --tool claude-code
# 或者只装某一个部门
cp engineering/*.md ~/.claude/agents/

For other tools, a conversion script generates compatible files, followed by an interactive installer that detects installed tools:

# Step 1: 为所有支持的工具生成集成文件
./scripts/convert.sh
# Step 2: 交互式安装，自动识别你装了哪些工具
./scripts/install.sh
# 或者直接指定一个工具
./scripts/install.sh --tool antigravity
./scripts/install.sh --tool gemini-cli
./scripts/install.sh --tool opencode
./scripts/install.sh --tool copilot
./scripts/install.sh --tool openclaw
./scripts/install.sh --tool cursor
./scripts/install.sh --tool aider
./scripts/install.sh --tool windsurf
./scripts/install.sh --tool kimi

The scripts support Claude Code, GitHub Copilot, Antigravity, Gemini CLI, OpenCode, OpenClaw, Cursor, Aider, Windsurf, and Kimi Code.

Usage

After installation, an agent is activated with a single command, for example:

Hey Claude, activate Frontend Developer mode and help me build a React component

Or in Chinese:

帮我激活「Filament 优化专家」，把这个后台资源页重构一下，要更快更干净

Activated agents follow their predefined persona and SOP. The “Threat Detection Engineer” first asks about detection tier, ATT&CK mapping, and SIEM status before providing code.

Empirical Evaluation

Codebase Onboarding Engineer : reads an unfamiliar codebase, maps modules, and explains them in plain language without modifying files.

Code Reviewer : produces structured reviews categorized by security, maintainability, and performance, improving over default Claude Code reviews.

Git Workflow Master : guides through branch strategy, Conventional Commits, and history cleanup.

Reality Checker : when asked to verify a release, generates an evidence checklist and prompts for missing items, demonstrating strict release gating.

Limitations

All agents are defined in English; Chinese scenarios require additional prompting to adjust tone.

Depth of specialized roles (e.g., smart‑contract or embedded firmware) depends on the amount of context supplied.

Agent‑to‑agent collaboration must be orchestrated manually; the repository does not provide an out‑of‑the‑box multi‑agent system.

Installing the full set of 170+ agents can confuse Claude Code’s selection mechanism; selecting 5‑10 relevant agents is recommended.

Illustrative Images