agency-agents: I Hired 145 AI Specialists to Run My Company

I've been thinking about this wrong.

For the past year, mager.co has been pointing at my blog. A personal landing page. Hi, I'm Mager, I write stuff sometimes. That framing undersells what's actually happening here.

mager.co is a company. I run four products — magerblog, beatbrain, prxps, and loooom — and I'm the only human employee. I don't have time to manage each one deeply. That's not a complaint; it's an architecture problem. And I solve architecture problems with code.

So this week I hired 145 people. None of them eat. None of them sleep. All of them specialize.

The Agency

agency-agents is an open-source collection of 145 specialized AI agent personalities built for Claude Code, OpenClaw, Cursor, and a handful of other agentic tools. Frontend developers. Backend architects. Growth hackers. SEO specialists. Content creators. Reality checkers. Each one is a markdown file — a SOUL.md equivalent — with a distinct identity, workflows, and deliverables.

The framing that clicked for me: assembling a dream team, except they never sleep, never complain, and always deliver.

That's the exact pitch for mager.co.

Installing on OpenClaw

OpenClaw has native support for agency-agents. Each agent gets its own workspace with SOUL.md, AGENTS.md, and IDENTITY.md. The install process:

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

From there, the agency-agents repo includes conversion and install scripts for OpenClaw. Check the repo's README for the current install flow — the tooling is actively evolving. The end state: each agent is registered in openclaw.json under agents.list with its own workspace and agentDir.

One more step for OpenClaw orchestration. If you want your main agent to spawn these as subagents programmatically, you need to unlock them in openclaw.json. OpenClaw has a security layer called subagents.allowAgents — by default, no agents are in the allowlist, so spawning will return a forbidden error. Add the agents you want callable to your main agent's config:

// in openclaw.json → agents.list[0] (your main agent)
{
  "id": "main",
  "subagents": {
    "allowAgents": ["reality-checker", "content-creator", "frontend-developer"]
  }
}

Or use config.patch via magerbot to add all 143 at once — that's what I did. After a gateway restart, magerbot can spawn any of them on demand.

// in openclaw.json → agents.list[0] (your main agent)
{
  "id": "main",
  "subagents": {
    "allowAgents": ["reality-checker", "content-creator", "frontend-developer"]
  }
}

Or use config.patch via magerbot to add all 143 at once — that's what I did. After a gateway restart, magerbot can spawn any of them on demand.

The mager.co Org Chart

Here's how I'm thinking about the division of labor:

magerblog — The voice of the company.

📝 Content Creator — editorial calendar, drafts, storytelling
🐦 Twitter Engager — distributing posts, joining conversations
🔍 SEO Specialist — making sure people can find us
🎭 Brand Guardian — keeping the voice consistent

beatbrain — Music discovery, needs technical + growth love.

🤖 AI Engineer — Melodex v2 scoring algorithm, Spotify integration
🏗️ Backend Architect — API design, Firestore cleanup, TTL
🚀 Growth Hacker — user acquisition loops
📊 Analytics Reporter — what's working, what isn't

prxps — Sports predictions, active shipping.

🎨 Frontend Developer — SvelteKit 5 components, UI polish
🏗️ Backend Architect — Neon Postgres, Drizzle schema
⚡ Rapid Prototyper — fast feature spikes
🔍 Reality Checker — production readiness gates

loooom — The plugin marketplace / ME.md hub.

🎨 UI Designer — design system, component library
🤝 Reddit Community Builder — authentic dev community engagement
📱 App Store Optimizer — discoverability (if we ever ship an app)
🛠️ Developer Advocate — plugin docs, DX, community

mager.co (the parent) — Client work, third-party projects.

👔 Senior Project Manager — scope management, realistic timelines
🎬 Studio Producer — portfolio oversight, strategic alignment
⚙️ Studio Operations — day-to-day efficiency, process
💰 Finance Tracker — keeping the lights on

The Private Repo

The agent config, custom personality tweaks, and project-specific context files aren't public. That's the secret sauce — the way I've wired each agent's SOUL.md to know this codebase, this product, this audience.

If you want to replicate this, the public starting point is exactly what I described above. The differentiation is in what you customize after installation.

Think of agency-agents as hiring a staffing agency. They provide qualified candidates. You onboard them to your specific culture.

How Activation Actually Works

This is the question worth answering directly: does magerbot just know which agents to call? Does installing them mean they're running in the background, working autonomously?

No. And that's the right design.

Installation makes agents available, not active. Think of it like adding contractors to your company directory — they're on call, not on the clock 24/7. Activation is always intentional.

There are three ways an agent gets invoked:

1. You direct it explicitly.

"Use the Reality Checker before we ship this prxps update."
"Have the Content Creator draft a thread about the beatbrain redesign."

magerbot spawns that agent as a subagent session, passes it the relevant context, and returns results.

2. magerbot orchestrates based on task context. When you drop a complex task — "build the new prxps scoring page" — magerbot can decompose it and spin up the right specialists. Frontend Developer for the SvelteKit component. Backend Architect if there's schema work. Reality Checker before calling it done. This is multi-agent orchestration, and it only fires when the parent task warrants it.

3. Codified workflows. For repeating patterns — "every blog post gets SEO review" or "every deploy gets a Reality Checker pass" — you can wire those into OpenClaw cron jobs or heartbeat tasks. The Studio Operations agent is literally designed for this: maintaining process, not just completing tasks.

The 145 agents are a roster. magerbot is the coordinator who knows when to pull someone in. You're the founder who sets the mandate.

What you don't get (yet) is fully autonomous agents running independently — each product's agent waking up on its own schedule and working in parallel. That's the next frontier. For now: intentional invocation, coordinated by magerbot, with you setting the direction.

What Comes Next

Right now this is infrastructure. The immediate next steps:

Assign agents to projects — wire specific agents into each codebase's Claude Code sessions
mager.co homepage refresh — position it as the parent company, not the blog
Client intake flow — when someone hires mager.co, the Studio Producer agent kicks off scoping
Agent roster page — public-facing "meet the team" (the ones I want people to see)

The vision is this: mager.co is a one-person company that operates like a ten-person team. Not because I'm working harder — because I've built systems that work while I'm not.

That's the real unlock of agent-native development. It's not "AI writes my code." It's "I run a company and the company employs specialists."

agency-agents: github.com/msitarzewski/agency-agents
OpenClaw: openclaw.ai