• Each agent is a detailed markdown personality file with identity, mission, workflow, and deliverables
• Organized into 11 departments mirroring a real company structure
• Not generic prompts -- agents include real code templates, success metrics, and non-negotiable rules
• Born from Reddit discussions and months of community iteration

agency-agents/
├── engineering/        # 23 agents
├── design/             # 8 agents
├── marketing/          # 26 agents
├── sales/              # 8 agents
├── product/            # 5 agents
├── project-management/ # 6 agents
├── testing/            # 8 agents
├── support/            # 6 agents
├── spatial-computing/  # 6 agents
├── specialized/        # 22 agents
├── academic/           # research
├── strategy/           # strategy
├── scripts/            # tooling
│   ├── convert.sh
│   ├── install.sh
│   └── lint-agents.sh
└── examples/           # usage demos

Naming Convention

Every agent file follows the pattern:

{division}-{role-name}.md

# Examples:
engineering-frontend-developer.md
design-ux-researcher.md
marketing-seo-specialist.md
specialized-agents-orchestrator.md

# Frontend Developer Agent
Name: Frontend Developer
Emoji: 🖥️  Color: Cyan
Vibe: "Pixel-perfect precision"

## Missions
1. Editor Integration Engineering
2. Web Application Development
3. Performance & UX Optimization
4. Code Quality & Scalability

## Critical Rules
- Lighthouse > 90 always
- WCAG 2.1 AA compliance
- LCP < 2.5s, FID < 100ms

## Communication Style
"Reduced render time by 80%"

Four Mission Areas

• Editor Integration -- WebSocket/RPC bridges, sub-150ms latency navigation
• Web App Development -- Responsive apps, component libraries, design systems, API integration
• Performance Optimization -- Core Web Vitals (LCP <2.5s, FID <100ms, CLS <0.1), code splitting, PWAs
• Code Quality -- TypeScript, testing suites, error handling, maintainable architecture

Non-Negotiable Rules

# Performance-First
Lighthouse > 90 for all metrics
Optimize from project inception
Monitor Web Vitals continuously

# Accessibility
WCAG 2.1 AA compliance always
Proper ARIA labels
Semantic HTML required
Keyboard nav + screen reader

Success Metrics

• Page loads under 3s on 3G
• 80%+ component reusability
• Zero console errors in prod

Western Platforms

China Market Specialists

Design (8 Agents)

Paid Media (7 Agents)

Sales (8 Agents)

Product (5 Agents)

Testing (8 Agents)

• Evidence Collector -- Screenshot-based QA validation
• Reality Checker -- Comprehensive integration testing across the full app
• Test Results Analyzer -- Interprets test output, identifies patterns
• Performance Benchmarker -- Load testing, bottleneck identification
• API Tester -- Endpoint validation, contract testing
• Tool Evaluator -- Assesses third-party tools and libraries
• Workflow Optimizer -- Process improvement and automation
• Accessibility Auditor -- WCAG compliance validation

Support (6 Agents)

• Support Responder -- Customer-facing ticket resolution
• Analytics Reporter -- Dashboards, KPI tracking
• Finance Tracker -- Budget monitoring, cost analysis
• Infrastructure Maintainer -- Server health, uptime
• Legal Compliance -- Policy checks, GDPR, SOC2
• Executive Summary Generator -- C-suite reporting

Spatial Computing (6 Agents)

• XR Interface Architect -- Cross-platform spatial UI design
• macOS Spatial/Metal Engineer -- Apple GPU and spatial frameworks
• XR Immersive Developer -- VR/AR experience development
• XR Cockpit Interaction -- Dashboard UX for spatial environments
• visionOS Spatial Engineer -- Apple Vision Pro apps
• Terminal Integration Specialist -- CLI tools for spatial workflows

Specialized (22 Agents) -- Highlights

• Agents Orchestrator -- The meta-agent that coordinates all others
• MCP Builder -- Builds Model Context Protocol servers
• Blockchain Security Auditor -- Smart contract security review
• ZK Steward -- Zero-knowledge proof systems
• Identity Graph Operator -- User identity resolution
• Document Generator -- Automated report creation
• Compliance Auditor -- Regulatory compliance
• Healthcare Marketing Compliance -- HIPAA-aware marketing

Four Sequential Phases

Quality Gates

# Decision Logic
if task passes QA:
    mark_complete()
    advance_to_next()
    reset_retry_counter()

elif task fails QA:
    increment_attempts()
    if attempts < 3:
        loop_back_with_feedback()
    else:
        escalate()
        continue_pipeline()

# Clone the repo
git clone https://github.com/\
msitarzewski/agency-agents.git
cd agency-agents

# Option A: Install for Claude Code
./scripts/install.sh
# Copies to ~/.claude/agents/

# Option B: Convert for other tools
./scripts/convert.sh
# Generates integration files

# Option C: Target a specific tool
./scripts/install.sh --tool cursor
./scripts/install.sh --tool windsurf
./scripts/install.sh --tool aider

# Validate agent files
./scripts/lint-agents.sh

Pattern 1: Personality-Driven

Each agent has a distinct communication style. The "Whimsy Injector" speaks differently from the "Security Engineer." This helps the LLM stay in character.

Pattern 2: Workflow-First

Not "you are an expert in X." Instead: "Step 1: do this. Step 2: check that. Step 3: deliver this artifact." Concrete, reproducible processes.

Pattern 3: Guardrails Built In

"Critical Rules" sections define what the agent must never violate. These act as hard constraints that prevent the LLM from drifting.

Pattern 4: Measurable Success

Every agent defines success metrics. Not "do a good job" but "Lighthouse > 90, LCP < 2.5s, zero console errors."

Pattern 5: Real Deliverables

Agent files include actual code templates, report formats, and output examples. The LLM knows exactly what "done" looks like.

Marketing Agents (China)

Engineering Agents (China)

• WeChat Mini Program Developer -- Builds apps within the WeChat super-app ecosystem
• Feishu Integration Developer -- Integrates with ByteDance's enterprise collaboration tool (China's Slack)

Commerce Agents

• China E-Commerce Operator -- Tmall, JD.com, Pinduoduo store management
• Cross-Border E-Commerce -- International selling into China
• Livestream Commerce Coach -- Live-selling on Douyin/Taobao Live
• Private Domain Operator -- WeChat group and CRM strategies

Strengths

• Incredible breadth -- 120+ agents covering nearly every business function
• Battle-tested -- Born from Reddit discussions and real-world usage
• Platform-agnostic -- Markdown files work with any AI tool
• Quality over vibes -- Real code, real metrics, real workflows
• Community-driven -- MIT license, active contributions
• Global perspective -- Deep China market coverage is rare

Limitations

• No runtime -- These are prompts, not executable code. You need an AI tool to run them.
• Orchestration is conceptual -- The orchestrator agent describes a pattern, but doesn't enforce it in code.
• Agent quality varies -- Community contributions mean some agents are more polished than others.
• Context window pressure -- Loading a full agent file plus your task can consume significant context.
• No memory -- Agents don't persist state between sessions unless the platform provides it.

Do

• Start with one agent -- Pick the role closest to your current task and use it for a week
• Customize the persona -- Remove missions that don't apply. Add your tech stack to the rules.
• Use the workflow section -- The step-by-step processes are often more valuable than the identity
• Combine 2-3 agents max -- Orchestrate a frontend dev + QA agent for a focused dev loop
• Check success metrics -- They're great checklists even without AI

Don't

• Load all 120+ agents -- Your context window will explode and quality will plummet
• Use them verbatim -- Adapt to your specific project, tech stack, and constraints
• Expect autonomous orchestration -- The orchestrator pattern requires manual setup per platform
• Skip the critical rules -- They're the most valuable part for preventing LLM drift
• Ignore the linter -- Run lint-agents.sh if you modify or create agents

# my-division/my-agent.md

## Identity
Name:    Your Agent Name
Emoji:   🔧
Color:   Pick a color
Vibe:    "One-line personality"

## Personality
- 3-5 core traits
- How they communicate
- What they remember

## Primary Missions
1. First responsibility area
2. Second responsibility area
3. Third responsibility area

## Critical Rules
- Non-negotiable constraint 1
- Non-negotiable constraint 2

## Workflow
Step 1: Setup → Step 2: Execute
→ Step 3: Validate → Step 4: Deliver

## Deliverables
# Include actual templates

## Success Metrics
- Measurable outcome 1
- Measurable outcome 2

Checklist

• File named {division}-{role}.md
• Identity section with emoji, color, and vibe
• At least 2 primary missions
• At least 2 non-negotiable critical rules
• Step-by-step workflow (not just "be an expert")
• At least one deliverable template with real code
• Quantifiable success metrics