TL;DR — Key Takeaways
  • Agency Agents is a free, open-source library of 230+ AI agent personas you install into Claude Code, Cursor, and 6 other tools.
  • Each persona gives your AI a specific domain identity — like "Senior Backend Architect" — with defined workflows, deliverables, and anti-hallucination rules.
  • The project hit 126,000+ GitHub stars within weeks — one of the fastest-growing open-source repos in 2026.
  • It's different from Claude's native sub-agents — those are an architectural execution feature; Agency Agents are persona definitions.
  • MIT licensed. No cost. No account needed. Install in under 2 minutes →

What is Agency Agents? The Complete Guide (2026)

📅 Updated July 2026⏱️ 8 min readView on GitHub ↗

The Short Version

If you've been using Claude Code, Cursor, or GitHub Copilot and getting generic AI responses — Agency Agents fixes that.

Instead of your AI acting as a general-purpose assistant, Agency Agents gives it a specific professional identity: Senior Backend Architect, Security Auditor, UX Researcher, or any of the 230+ other roles in the library.

Each agent persona is a carefully engineered Markdown file that defines the AI's mission, working methods, communication style, deliverables, success criteria, and failure patterns to avoid.

Why Developers Are Obsessed With It

The project hit 126,000+ GitHub stars faster than almost any developer tool in 2026. That velocity happened for a reason: Agency Agents solves a real problem that every Claude Code and Cursor user experiences.

The Problem Without Agency Agents

When you ask a generic AI to “help with the database layer,” it tries to do everything at once — architecture decisions, performance optimizations, security concerns, ORM choices — and ends up being mediocre at all of them while hallucinating half the specifics.

The Solution With Agency Agents

You activate the Data Engineer agent. It now operates with a specific mission (data pipelines, performance, reliability), defined deliverables (schema designs, migration scripts), and clear anti-patterns to avoid. The output quality jumps immediately.

How Agency Agents Works (Technically)

The mechanism is elegant in its simplicity. Each agent persona is a Markdown file stored in your AI tool's agents directory:

# Claude Code location
~/.claude/agents/
├── engineering/
│ ├── backend-architect.md
│ ├── frontend-developer.md
│ └── security-engineer.md
├── design/
├── marketing/
└── ... 16 divisions total

When you invoke an agent — for example, by saying “Use the Security Engineer agent to review this code” — Claude Code reads the corresponding Markdown file and uses it as a structured system prompt for that interaction.

The result: the AI now operates with the context, goals, constraints, and success metrics of a senior security professional, not a generalist.

The 16 Divisions — Full Breakdown

As of July 2026, Agency Agents includes 230+ personas across 16 functional divisions. The widely cited “147 agents” figure is outdated — the project grew significantly after its initial viral moment.

Engineering45+

Backend Architect, SRE, DevOps, Security

Design18+

UX Researcher, Visual Designer, Spatial Computing

Marketing22+

SEO Specialist, Brand Strategist, Content Lead

Product16+

Product Manager, Growth Hacker, Analytics

Security12+

Threat Detection, Pen Tester, Compliance

Strategy14+

Business Strategist, Competitive Intel

Game Dev15+

Game Designer, Level Architect, AI NPC

Finance11+

Financial Analyst, Revenue Modeler

Browse the complete agent catalog with all 16 divisions →

Agency Agents vs Claude Native Sub-Agents

This is the most common point of confusion, and it's worth addressing clearly.

FeatureAgency AgentsClaude Sub-Agents
What it isLibrary of persona definitionsBuilt-in task delegation architecture
FunctionDefines who the AI is and how it behavesDefines how tasks are spawned and delegated
PlatformWorks across Claude Code, Cursor, Copilot, 5+ moreSpecific to Claude Code CLI
ImplementationMarkdown files you installAutomatic — Claude spawns them as needed
Use together?✅ Yes — personas define the agent, sub-agents run it✅ Yes — complementary, not competing

Read the full comparison: Agency Agents vs Claude Sub-Agents →

Supported Tools (July 2026)

Agency Agents works with 8 AI coding tools, with deterministic rendering for each:

Claude CodeCodexCursorGitHub CopilotGemini CLIQwen CodeopencodeOsaurus

More tools (Antigravity, Aider, Windsurf, OpenClaw, Kimi) are on the roadmap but not yet exposed as first-class installs. See all install guides →

Ready to install Agency Agents?

Setup takes under 2 minutes. Pick your tool below.

Frequently Asked Questions

Is Agency Agents the same as Claude's built-in sub-agents?

No. Agency Agents is a library of persona definitions (Markdown files) that define who your AI is. Claude's native sub-agents are an architectural feature that lets Claude spawn isolated helper instances for parallel tasks. They are complementary — use Agency Agents to define the persona, use native sub-agents to orchestrate the execution.

Does Agency Agents require internet access to work?

No. The personas are Markdown files stored locally in your ~/.claude/agents/ directory (or equivalent). They work entirely offline. The desktop app (agencyagents.app) also supports offline-first mode with a bundled catalog.

How is Agency Agents different from just writing a prompt?

Agency Agents personas are structured definitions with identity, mission, working methods, deliverables, success criteria, and anti-patterns. A typical persona is 200-400 lines of carefully engineered instructions that reduce hallucinations and keep the AI on-task better than a casual prompt.

Can I use Agency Agents with VS Code or Windsurf?

Cursor (a VS Code fork) is fully supported. Pure VS Code with GitHub Copilot is also supported. Windsurf, Aider, and a few others are listed on the roadmap but not yet fully exposed as first-class installs in the desktop app.

How often is the catalog updated?

The GitHub repo is actively maintained by msitarzewski and community contributors. New agents and divisions are added frequently. The desktop app auto-updates the catalog from the upstream repository.

What does '147 agents' mean? I see different numbers everywhere.

147 was the agent count when the repo first went viral in June 2026. The project has grown significantly since — as of July 2026, there are 230+ agents across 16 divisions. The number you see depends on when the article was written. Always check the live GitHub repo for the current count.