🎭 The Agency: 232 Specialized AI Agents to Transform Your Workflow

Core question this article answers: “As a developer or product leader, can I have a ready‑to‑use ‘virtual team’ of experts across every domain without actually hiring dozens of people?”

Yes. And that team already exists. It is not another set of generic prompts – it is a carefully crafted collection of AI agents with distinct personalities, hardcoded workflows, and measurable deliverables. Each agent is deeply specialised: frontend development, security auditing, Reddit community building, supply chain strategy – they have a voice, a process, and proven output. This article walks you through the open‑source project behind these agents, how to install them, how to combine them, and why they work differently from simple role‑playing prompts.

🚀 Origin: From a Reddit Thread to 232 Production‑Ready Agents

Core question of this section: “Where did this agent system come from, and what makes it different from the ‘act as a…’ prompts I see everywhere?”

In late 2023 a Reddit thread about making AI truly work like a specialist gained unexpected traction. Many people noticed that simple role‑playing prompts produce vague, shallow answers – not runnable code, not actionable design systems, not measurable marketing plans. So the project maintainer started writing, testing, and iterating on agent files that have fixed identities, enforced workflows, and explicit deliverables. After months of real‑world usage, the repository now contains 232 agents across 16 divisions, with ready‑made integrations for 12 AI coding tools: Claude Code, Cursor, GitHub Copilot, OpenCode, Windsurf, Aider, Antigravity, Gemini CLI, OpenClaw, Qwen Code, Kimi Code, and Codex.

How “The Agency” Differs from Ordinary Prompts

Every agent follows a fixed template: Identity & Memory → Core Mission → Critical Rules → Technical Deliverables (with code examples) → Workflow Steps → Success Metrics. This ensures the AI produces high‑quality, predictable output every time you invoke the agent.

⚡ Quick Start: Get Your Expert Team Running in Three Minutes

Core question of this section: “I don’t want to mess with config files. What’s the fastest way to use these agents locally?”

The project provides two shell scripts: convert.sh and install.sh. convert.sh transforms the 232 source Markdown agents into the format required by each target tool (Cursor’s .mdc rules, Aider’s CONVENTIONS.md, Antigravity’s SKILL.md, etc.). install.sh auto‑detects which tools are installed on your machine and lets you interactively choose which agents to install for which tools.

One‑Command Installation (recommended)

# Clone the repository
git clone <repository-url>
cd agency-agents

# Generate integration files for all supported tools (parallel is faster)
./scripts/convert.sh --parallel

# Interactive installer – auto‑detects installed tools, shows a checklist
./scripts/install.sh

After installation, you can invoke agents directly in your tool. For example, in Claude Code:

“Hey Claude, activate Frontend Developer mode and help me optimise this React component’s performance.”

In Cursor you can write:

“Please review this API code using the @security-engineer rules.”

Install Only the Divisions You Need

If you don’t want all 232 agents (e.g. you only care about engineering and security):

# Install only engineering and security divisions to Claude Code
./scripts/install.sh --tool claude-code --division engineering,security

# Install only specific agents to Cursor
./scripts/install.sh --tool cursor --agent frontend-developer,code-reviewer

A quick reflection: Many teams try to “use everything” when adopting AI, which bloats context and slows down responses. The project’s ability to install a subset (by division or by individual agent) is a practical design choice. It saves tokens and keeps the AI focused. That “minimal viable agent set” approach is worth adopting in any AI‑powered workflow.

🧩 The Agency Roster: 16 Divisions, 232 Roles at a Glance

Core question of this section: “Besides developers, what other roles are covered? Is there an agent for my specific business function?”

Below is a high‑level overview of all divisions and a few representative agents from each. Scan the tables to find what matches your needs.

💻 Engineering Division

From embedded firmware to cloud‑native architecture.

Real‑world scenario: When you need to break a monolithic PHP app into microservices and move to the cloud, you can simultaneously activate “Backend Architect” (to design API boundaries), “DevOps Automator” (to write Terraform scripts), and “Security Architect” (to enforce zero trust). The agents work in parallel and deliver a single, actionable migration plan.

🎨 Design Division

More than visuals – brand, inclusivity, and delight.

• UI Designer – component libraries, design systems
• UX Researcher – user testing, behaviour analysis
• Brand Guardian – brand identity, positioning
• Whimsy Injector – micro‑interactions, Easter eggs, brand personality
• Inclusive Visuals Specialist – bias mitigation in AI‑generated imagery
• Persona Walkthrough Specialist – cognitive walkthroughs based on user personas

Author’s insight: I was particularly struck by the “Whimsy Injector” agent. It does not treat “fun” as an add‑on – every playful element must serve a functional or emotional purpose. That is exactly what many products miss: a well‑placed celebration animation can reduce task completion anxiety by 40% (as noted in the agent’s success metrics). This is a design philosophy worth stealing.

💰 Paid Media Division

Turning ad spend into measurable outcomes – from audit to creative.

• PPC Campaign Strategist – Google/Microsoft Ads account architecture
• Search Query Analyst – negative keywords, intent mapping
• Paid Media Auditor – 200+ point account checks
• Tracking & Measurement Specialist – GTM, GA4, CAPI
• Ad Creative Strategist – RSA copy, Meta creatives, testing programmes

Real‑world case (from the project documentation): A team took over an underperforming ad account. In 30 days they sequentially used “Paid Media Auditor” (full assessment) → “Tracking & Measurement Specialist” (validate conversion data) → “PPC Campaign Strategist” (redesign account structure) → “Search Query Analyst” (clean wasted spend) → “Ad Creative Strategist” (refresh all copy). The result: waste eliminated, structure optimised, creative renewed – all led by AI agents, with humans only doing final approval.

📢 Marketing Division

Covers domestic and international social platforms and content ecosystems.

• Growth Hacker – viral loops, experiment design
• Content Creator – multi‑platform editorial calendar
• Reddit Community Builder – authentic engagement, value‑first
• TikTok Strategist – algorithm optimisation, viral content
• Xiaohongshu Specialist – lifestyle content, trend‑driven strategy
• Baidu SEO Specialist – China search optimisation, ICP compliance
• Bilibili Content Strategist – danmaku culture, UP‑holder growth
• Podcast Strategist – Chinese podcast market, platform optimisation

“You’re not marketing on Reddit – you’re becoming a valued community member who happens to represent a brand.” – Reddit Community Builder (agent self‑description)

Other 12 Divisions – Quick Reference

Reflection: Looking at this list, I realise how far AI capabilities have gone beyond “writing code” or “writing copy”. An agent like “Cultural Intelligence Strategist” is expected to understand cultural exclusion in global UX. “China Government Digital Presales Consultant” needs to know the tender process and digital transformation proposals. This is simulating high‑context, senior professional roles. A startup could realistically use these 232 agents to replace an initial team of 20 people – not by replacing humans entirely, but by letting each human “command” dozens of AI experts.

🎯 Four Real‑World Scenarios: From MVP to Enterprise Delivery

Core question of this section: “How do I combine these agents in practice? Are there ready‑made team compositions I can copy?”

Scenario 1: Building a Startup MVP

Your AI team:

1. Frontend Developer – builds the React app
2. Backend Architect – designs API and database
3. Growth Hacker – plans user acquisition
4. Rapid Prototyper – iterates fast
5. Reality Checker – quality gate before launch

Workflow: Backend Architect outputs data models and API specs → Frontend Developer builds components in parallel → Rapid Prototyper delivers a demo every two days → Growth Hacker designs cold‑start campaigns → Reality Checker demands screenshot/log evidence before signing off. No human coordination needed – just assign agents in separate conversations or sequentially.

Scenario 2: Enterprise Feature Development

Team composition:

• Senior Project Manager – breaks down specs into tasks
• Senior Developer (Laravel/Livewire) – implements complex logic
• UI Designer – delivers design‑system‑ready components
• Experiment Tracker – designs A/B tests
• Evidence Collector – screenshot‑level UI testing
• Reality Checker – production readiness certification

Deliverables: a task‑broken project board, a PR with design‑system components, an A/B test plan, a test report with visual evidence, and a “go live” certificate.

Scenario 3: Omnichannel Marketing Campaign

Agents used:

• Content Creator – master editorial calendar
• Twitter Engager – real‑time engagement
• Instagram Curator – visual storytelling
• Reddit Community Builder – authentic interaction
• Analytics Reporter – cross‑platform data integration

Value: Each platform is handled by the agent that understands its algorithm best, while brand voice stays consistent. The marketing manager only reviews final posts and dashboards – no more writing separate copies for each platform.

Scenario 4: Smart Campus Digital Twin (GIS + BIM + Drone)

Your AI team:

1. GIS Technical Consultant – define digital twin strategy (BIM buildings + GIS campus + IoT real‑time)
2. BIM/GIS Specialist – convert Revit models to GIS scene layers, indoor floor plans
3. Drone/Reality Mapper – fly the campus, generate orthomosaic and 3D mesh
4. Web GIS Developer – build interactive campus dashboard with MapLibre
5. 3D & Scene Developer – terrain, buildings, flyover tour
6. GeoAI/ML Engineer – extract building footprints and tree canopy from imagery
7. GIS QA Engineer – validate topology, CRS consistency

Result: a complete campus digital twin that fuses BIM details, drone reality capture, 3D visualisation, and web access – delivered by a pipeline of agents, compressing months of work into one week.

Author’s reflection: In the past, a digital twin project required hiring separate BIM engineers, GIS specialists, WebGL frontend developers, drone pilots, and AI annotation teams. Now one person who understands the business domain can sequentially activate these agents. The GIS division in this project contains every one of those roles, each with detailed deliverables (e.g. the “BIM/GIS Specialist” must output Revit‑to‑IFC conversion scripts and indoor mapping services). This granularity makes me believe that truly production‑ready AI workflows have arrived.

🔌 Twelve‑Tool Integration Deep Dive

Core question of this section: “I use Cursor / Copilot / Windsurf / Claude Code daily – can I use these agents with my existing tools, and how?”

The project natively supports the tools listed below, with one‑command generation and installation scripts. Here are the most common integrations.

Claude Code (native)

Agents are copied directly to ~/.claude/agents/ – no conversion needed. Just mention the agent name in your conversation.

Cursor

Each agent becomes an .mdc rule file in .cursor/rules/ of your project. Cursor auto‑detects them; you can also explicitly reference with @agent-name.

GitHub Copilot

Similar to Claude Code – agents go into ~/.github/agents/ and ~/.copilot/agents/.

Aider

All agents are merged into a single CONVENTIONS.md file in your project root. Aider reads it automatically. Usage: “Use the Frontend Developer agent to refactor this component.”

Windsurf

All agents become .windsurfrules in your project root. Reference agent names in Cascade.

OpenCode

Agents are placed in .opencode/agents/ (project‑scoped) or ~/.config/opencode/agents/ (global). Note: OpenCode currently silently drops any agent count exceeding ~119, so install a single division with --division to stay under the limit. The installer warns you.

Quick reference for other tools

Installation advice: Start small. Install only 1–2 agents for one tool and experience the workflow before expanding.

./scripts/install.sh --tool cursor --agent frontend-developer,code-reviewer

📈 Success Metrics & Quality Gates

Core question of this section: “How do I know if an agent’s output is good? Are there quantifiable standards?”

Every agent file includes built‑in Success Metrics. For example, the “Frontend Developer” agent’s metrics:

• First Contentful Paint (FCP) < 1.5s
• Interaction to Next Paint (INP) < 200ms
• Lighthouse performance score > 90
• Accessibility (a11y) compliance report provided

The “Reality Checker” agent goes further: any claim of “fixed” must be accompanied by screenshots, logs, or test results – otherwise it refuses to certify. This “evidence‑enforced” mechanism eliminates hallucinated fixes.

Reflection: This is one of the most engineering‑valuable designs in the project. Most AI prompts only tell the AI what to do, not what “done well” looks like. When you embed success metrics into the agent’s system prompt, the AI self‑evaluates before answering. I have seen teams use the “Reality Checker” as an automated PR gate – a merge cannot happen until this agent issues a “production ready” certificate backed by performance benchmarks and test coverage data. This is not hype; it is an automatable quality system.

❓ Frequently Asked Questions (FAQ)

1. Do I have to pay for these agents?
   No. The project is MIT licensed – free for commercial and personal use. You can sponsor the maintainer on GitHub if you wish.

2. Can I use these agents with Chinese prompts?
   Yes. The agent files are written in English, but you can prompt the AI in Chinese. The agents follow their internal English rules but will reply in your language. Community‑maintained Simplified Chinese, Japanese, and Korean translations already exist.

3. Can I modify or add my own agents?
   Absolutely. Fork the repository, create a new .md file following the existing template, and submit a PR. Contributions are welcome, especially real‑world success stories and improved success metrics.

4. Will the agents leak my code or business secrets?
   The agents are just text prompts that run inside your local AI tool (Claude Code, Cursor, etc.). Data upload depends on the privacy policy of the AI service you use. Avoid pasting secrets or customer PII directly into the conversation.

5. Will installing all 232 agents overload the AI’s context?
   Yes. The documentation explicitly advises not to install all agents into one tool. For Cursor, for example, you should only copy the .mdc files of the 2–3 agents relevant to your current project into .cursor/rules/.

6. Can these agents be used for non‑coding tasks?
   Definitely. Marketing, sales, product, security, and many other divisions produce strategic documents, compliance checklists, data analyses, and more. You can even copy an agent’s content into a plain ChatGPT session and start role‑playing.

7. Are the “success metrics” automatically checked?
   No. They are self‑assessment guidelines for the AI. You can, however, write your own scripts to automatically validate the code or documents the agent outputs.

8. My favourite tool isn’t on the support list. What can I do?
   The agents are plain Markdown. You can manually copy the relevant content into whatever prompt format your tool accepts. Or you can add a new conversion target to scripts/convert.sh and submit a PR.

📦 One‑Page Summary (Action Checklist)

• [ ] Clone the repository
• [ ] Run ./scripts/convert.sh --parallel to generate tool‑specific formats
• [ ] Run ./scripts/install.sh to interactively select tools and agents
• [ ] (Optional) Install only a specific division: --division engineering,security
• [ ] (Optional) Install only specific agents: --agent frontend-developer
• [ ] Invoke agents in your IDE/CLI, e.g. “Activate Code Reviewer agent to review my PR”
• [ ] Verify output against the agent’s built‑in success metrics
• [ ] Tweak or add new agents based on your experience and contribute back

Practical Summary: Reshape Your Workflow with an AI Expert Team

This article introduced an open‑source collection of 232 production‑grade AI agents, covering 16 divisions from engineering to GIS and game development. The key takeaways:

1. From role‑play to executable workflow – Each agent has a fixed identity, an enforced step‑by‑step process, explicit deliverable formats, and measurable success metrics. This yields far more stable output than generic prompts.
2. Install only what you need – The project supports granular installation by division, by individual agent, and by target tool. Best practice: install just 3–5 agents relevant to your current project.
3. Unified experience across tools – Whether you use Claude Code, Cursor, Aider, or Windsurf, the conversion scripts make the same agent knowledge portable. One asset, many platforms.

If you are looking for a practical, immediately usable way to bring AI expertise into your daily work – or you want to build your own library of specialist agents – this project is a solid starting point. Go ahead, let those 232 tireless experts start working for you.