title: "You Installed 200 AI Experts, But It Only Uses 3—Is the 128K Star Agent Library Worth It?" description: "The GitHub 128K star project agency-agents includes 200+ professional AI Agents covering 20+ departments like engineering, design, marketing, and security. But more installed doesn't mean better use—context bloat, uneven quality, role drift—this article clarifies the costs." tags: ["AI Agent", "agency-agents", "Claude Code", "Cursor", "Multi-Agent Collaboration"]
You Installed 200 AI Experts, But It Only Uses 3—Is the 128K Star Agent Library Worth It?
I admit, the moment I cloned agency-agents, I felt like I had hired an entire company. Frontend development, backend architecture, security audit, marketing growth, and even a "WeChat Mini Program Developer"—over 200 people, not a cent spent.
Then I realized I couldn't even get them to work.
It wasn't the tool's problem; it was mine. With 200 Agents stuffed into .claude/agents/, Claude Code reads every personality file on startup—context consumed before any work begins. Worse, most Agents I never need—my project doesn't require game development, GIS geoinformation, or someone to design Roblox avatars.
A 128K star project, but less than 10 might be useful to you.
From a Reddit Post to 128K Stars
Agency-agents had humble beginnings. In late 2025, someone on Reddit posted a discussion about whether AI could truly specialize in a domain rather than giving a generic answer every time. Project author Michael Sitarzewski followed this thread and began creating structured expert personality files for AI coding tools.
Each Agent isn't just a one-liner "Act as a developer" prompt—it's a complete Markdown file containing:
- Identity & Personality: Not a vague "expert" but a role with clear professional boundaries and communication style
- Core Mission: Do one thing to perfection
- Deliverables: Specific code, documents, processes—not empty advice
- Success Metrics: Quantifiable quality standards
- Workflow: Complete steps from receiving a task to delivery
This "structured personality" approach improved output quality by about 70% compared to generic prompts—based on a controlled test by YouTube channel FuturMinds using the same task and model.
The project expanded from a few engineering Agents to today's 20+ departments and 200+ Agents, with 367 community commits. Now there's a native desktop app supporting macOS, Linux, Windows, one-click install to 15+ tools like Claude Code, Cursor, Codex.
20 Departments, 200+ Experts—Don't Install Everything Yet
Agency-agents organizes Agents by department, each solving a set of problems:
You'll likely use:
- Engineering (30 people): Frontend, backend, DevOps, code review, database optimization—core
- Design (8 people): UI design, UX research, brand guardian, visual storytelling
- Product (5 people): Sprint planning, trend research, feedback synthesis, behavior nudging
- Security (11 people): Security architecture, penetration testing, cloud security, threat intelligence
You may not use:
- Game Development: Unity, Unreal, Godot, Roblox-specific Agents
- GIS Geoinformation: Spatial data, drone mapping, 3D scenes
- Spatial Computing: XR, Vision Pro, WebXR
- Academics: Anthropology, geography, narratology, psychology—for world-building
China market specific:
- Baidu SEO expert, Douyin strategist, Bilibili content strategist
- Xiaohongshu operator, Kuaishou strategist, private domain operator
- Multi-platform one-click publish (Zhihu, Xiaohongshu, CSDN, Bilibili, WeChat Official Account, Juejin)
- Feishu integration development, WeChat Mini Program development
Install in 3 Minutes, But Don't Install Everything
# Recommended: Install only the departments you need
./scripts/install.sh --tool claude-code --division engineering,security
# Or install only specific Agents
./scripts/install.sh --tool cursor --agent frontend-developer,code-reviewer
# macOS can also use Homebrew for the desktop app
brew install --cask msitarzewski/agency-agents/agency-agents
Why selective install? Three reasons:
Context Bloat: Claude Code reads all files in the agents directory on startup. 200 Markdown files, each 2-5KB, means 400KB-1MB of context eaten. Your effective window shrinks.
Role Conflict: Install Frontend Developer and Senior Developer. When you ask Claude to write frontend code, it may not know which role to follow, causing chaotic output styles.
OpenCode Hard Limit: OpenCode runtime registers at most ~119 Agents; extras are silently dropped. Full install means half your Agents disappear.
Cost Awareness: 128K Stars ≠ 128K Approval
The community's view of agency-agents isn't unanimous praise.
Uneven quality is the biggest issue. Data Science Dojo's review bluntly said: Code Reviewer is usable as a start, but AI Engineer is "buzzword soup." Among 200+ Agents, some are polished, others are clearly community PR filler—format aligned, but professional depth lacking.
Role drift still exists. Structured personalities reduce AI output inconsistency but don't eliminate it. When you ask a Backend Architect a frontend-related question, it won't say "That's not my domain"—it will try to answer, then give a poor frontend solution with a backend mindset.
Not a true multi-agent system. Agency-agents provides "personality switching," not "multi-agent collaboration." You can't have Security Architect automatically review Backend Architect's output. For real multi-agent workflows, you need to build the agent-to-agent communication and coordination yourself—agency-agents doesn't help there.
Maintenance costs are underestimated. The project updates frequently (367 commits). Each update may change Agent behavior. If you've built a workflow based on a specific Agent's output, you may need to re-adapt after an update.
Who Should Use It and Who Shouldn't
Suitable for you if:
- You're an indie developer or solo team needing multi-role coverage
- You use Claude Code or Cursor and want to quickly activate a specialized role
- You only need 5-10 core Agents, not the full set
- You're willing to spend time testing which Agents actually improve output quality
Not suitable for you if:
- You expect AI to automatically become an all-powerful team after install (it won't)
- You need true multi-agent automated collaboration (that's not what this project solves)
- Your context window is already tight (full install will make it worse)
- You demand every Agent to be production-grade quality (reality is uneven)
Trend Judgment: The Next Step of Agent Personalization
Agency-agents' real value isn't the 200 Markdown files—it validates a direction: AI's competitive dimension is shifting from 'model capability' to 'role definition.'
The same Claude Sonnet with a generic prompt vs structured Agent personality shows a 70% output quality gap—a gap models alone can't close. Future AI coding tools won't compete on who has a bigger model, but on who can define more precise roles, clearer delivery processes, and stricter success metrics.
Agency-agents is an early experiment in this direction. It's rough, uneven in quality, and full install has side effects—but it points the way for AI coding to evolve from "general assistant" to "specialized team."
Installing 3 Agents you actually use is better than installing 200 for show.
💡 Quick Start
- GitHub Repository: github.com/msitarzewski/agency-agents
- Desktop App: agencyagents.app (macOS / Linux / Windows)
- Homebrew:
brew install --cask msitarzewski/agency-agents/agency-agents- Recommended install:
./scripts/install.sh --tool claude-code --division engineering,security
暂无评论。