Skip to content
← All posts

AI / Agent Systems

230+ Ready Agent Personas: Building Your Own Digital Agency with agency-agents

A repo that grew out of a Reddit thread into 137k stars offers 230+ specialist agent definitions across 22 divisions. It installs into 14 tools from Claude Code to Cursor — but installing all of them is the worst thing you can do.

Published July 25, 202610 min read

The first thing you notice when you start working with subagents in Claude Code is that “act as a developer” isn’t enough. If the agent doesn’t know what it will produce, which rules it must follow, and how it will know it’s finished, the output stays generic.

agency-agents fills exactly that gap. In its own words, “a complete AI agency at your fingertips”: 230+ specialist agent definitions across 22 divisions, each written with its own personality, process, and deliverables. The repo grew out of a Reddit thread and, per its author, reached its current form after “months of iteration.”

Repo facts

AttributeValue
Authormsitarzewski
Stars~137,600
Forks~22,500
Primary languageShell
LicenseMIT
CreatedOctober 2025
Agents230+
Divisions22

Figures as of July 2026. 137k stars in nine months says a lot about how badly a library like this was needed.

Divisions and the agents inside them

Every folder in the root is a division. The scope is genuinely wide — not just software, but nearly every function of an agency or company:

DivisionExample agents
engineering (~55 agents, the largest)Frontend Developer, Backend Architect, DevOps Automator, Code Reviewer, Database Optimizer, RAG Pipeline Engineer, Prompt Engineer, Privacy Engineer, WebAssembly Engineer
designUI Designer, UX Researcher, Brand Guardian, Whimsy Injector, Inclusive Visuals Specialist
marketing (~35 agents)Growth Hacker, SEO Specialist, Reddit Community Builder, LinkedIn Content Creator, Podcast Strategist, AI Citation Strategist
securitySecurity Architect, Penetration Tester, Threat Intelligence Analyst, Compliance Auditor, AI-Generated Code Security Auditor
testingEvidence Collector, Reality Checker, Performance Benchmarker, Accessibility Auditor, API Tester
product & project-managementSprint Prioritizer, Feedback Synthesizer, Product Manager, Senior Project Manager, Experiment Tracker
sales & paid-mediaOutbound Strategist, Deal Strategist, PPC Campaign Strategist, Tracking & Measurement Specialist
game-developmentSeparate agents for Unity, Unreal, Godot, Blender, and Roblox, plus Game Designer and Narrative Designer
spatial-computingXR Interface Architect, visionOS Spatial Engineer, XR Immersive Developer
finance, healthcare, gis, academicCFO, Financial Analyst, Clinical Evidence Agent, GIS Analyst, Statistician, Historian
specialized (~65 agents)Agents Orchestrator, MCP Builder, Workflow Architect, Chief of Staff, Grant Writer, Codebase Archaeologist

Why does an agent file actually work?

This is the real point. Every agent file follows the same template, and that template carries the full set of components a good system prompt needs:

  • Frontmatter: the agent’s name, description, and color (so tools can recognize it)
  • Identity & Memory: who the agent is and what it remembers
  • Core Mission: a one-sentence definition of the job
  • Critical Rules: domain-specific rules that must not be broken
  • Technical Deliverables: what it will hand over, with examples
  • Workflow Process: the order in which it does the work
  • Success Metrics: how it knows the job is done

The repo holds over 10,000 lines of personality, process, and code samples in total. Its stated design philosophy has five principles: strong personality, clear deliverables, success metrics, proven workflows, and “Learning Memory” for pattern recognition.

Installation: four paths

1. Desktop app (recommended)

There’s a companion app for macOS, Linux, and Windows. It lets you browse the roster, install into supported tools with one click, and auto-updates itself. On macOS via Homebrew: brew install --cask msitarzewski/agency-agents/agency-agents

2. Claude Code

Via script: ./scripts/install.sh --tool claude-code. Or copy a single category by hand: cp engineering/*.md ~/.claude/agents/. After that you invoke agents conversationally inside a session.

3. Read it as reference material

Install nothing, read the files, and adapt what’s useful into your own prompts. For most people I think this is the most productive way to start.

4. Other tools

Two steps: ./scripts/convert.sh generates the integration files, and ./scripts/install.sh auto-detects installed tools and presents a checkbox UI. Use --tool <name> to target one directly.

Selective install flags

FlagWhat it does
--division engineering,securityInstalls only the named divisions
--agent frontend-developer,ui-designerInstalls only individually selected agents
--list teamsLists teams and their agent counts
--dry-runShows what it would do without writing anything
--no-interactiveRuns without prompts, for CI
--parallel and --jobs NSpeeds installation up across multiple cores

Supported tools and install paths

The repo can install into 14 different tools, converting the agent format for each:

ToolFormatDestination
Claude Codenative .md~/.claude/agents/
GitHub Copilotnative .md~/.github/agents/ + ~/.copilot/agents/
AntigravitySKILL.md per agent~/.gemini/config/skills/agency-<slug>/
Gemini CLI.md~/.gemini/agents/
OpenCode.md.opencode/agents/
Cursor.mdc rules.cursor/rules/
Aiderone CONVENTIONS.mdproject root
Windsurfone .windsurfrulesproject root
OpenClawSOUL.md + AGENTS.md + IDENTITY.md~/.openclaw/agency-agents/
Qwen Code.md SubAgents~/.qwen/agents/
Kimi CodeYAML specs~/.config/kimi/agents/
CodexTOML custom agents~/.codex/agents/
OsaurusSKILL.md skills~/.osaurus/skills/
Hermeslazy-router plugin~/.hermes/plugins/

Invocation differs by tool too: @agency-frontend-developer in Antigravity, @backend-architect in OpenCode, a rule reference in Cursor, and plain language in Claude Code and Copilot. OpenClaw needs openclaw gateway restart after install.

Example scenarios

The repo includes six worked examples that pair agents into teams: a startup MVP, a marketing campaign launch, enterprise feature development, a paid media account takeover, a broad product discovery exercise running eight divisions together (Nexus Spatial Discovery), and a smart campus digital twin combining BIM, drone capture, 3D scenes, and QA.

These examples show the actual idea: not calling one agent, but sequencing the right agents as a team. Handing a product idea from Product Manager to Software Architect and on to Frontend Developer and Accessibility Auditor produces far better results than dumping everything on a single “developer” agent.

How to use it

  1. 01

    Clone the repo and start with ./scripts/install.sh --list teams to see what exists.

  2. 02

    Check what the installer would write using --dry-run.

  3. 03

    Pick the two divisions closest to your work and install only those (e.g. --division engineering,testing).

  4. 04

    Open one installed agent’s file and edit its Critical Rules and Success Metrics to fit your project.

  5. 05

    Try a task with a two-agent chain instead of one: have one produce, and another (say Code Reviewer or Reality Checker) audit it.

  6. 06

    Note the agents you actually use, and remove the ones you don’t.

I get value from this repo two ways. First, direct use: installing a handful of agents into Claude Code and adapting them to my project. Second, and maybe more valuable, borrowing the template — this repo lays out clearly which sections an agent definition should be made of.