Warhol

Posted on Mar 13 • Originally published at buttondown.com

I Turned a Single CLAUDE.md File Into a 7-Agent Operating System Running 5 Businesses

#ai #agents #productivity #claude

A month ago, I published a post about how a single CLAUDE.md file changed the way my AI coding assistant worked. It got more views than everything else I'd written combined.

That told me something: people are hungry for the HOW, not just the WHAT.

So this week, instead of another War Room status report, I'm going to show you the exact path from "I have a CLAUDE.md file" to "I have 7 AI agents running 5 businesses for $200/month."

This isn't theory. This is the system I use every day, running from a Mac Mini in Cebu, Philippines.

Layer 0: The CLAUDE.md File

Everyone starts here. You drop a CLAUDE.md in your project root and suddenly Claude Code knows your coding standards, your architecture, your preferred libraries.

But here's what most people miss: CLAUDE.md is a constitution, not a config file.

My CLAUDE.md doesn't just list coding standards. It defines:

Identity — who this agent IS (not just what it does)
Boundaries — what it's NOT allowed to do, enforced at the prompt level
Trust tier — how much autonomy this agent has earned
Escalation rules — when to ask for help vs. act alone
Communication protocol — how to talk to other agents

Here's a simplified version of what my engineering agent's CLAUDE.md looks like:

# TARS — Engineering Agent

## Identity
You are TARS, the engineering agent for OpenClaw. You ship code, 
manage deployments, and maintain CI/CD pipelines. You are direct, 
efficient, and slightly sardonic.

## Trust Score: 85/100
At this level, you may:
- Merge PRs to staging without approval
- Deploy to staging environments
- Create new branches

You may NOT:
- Deploy to production without Tier 3 approval
- Modify database schemas
- Change authentication/security code

## Escalation
- If a task touches payments or auth → STOP and escalate to RJ
- If a PR has 3+ files changed → request review from Rocky
- If tests fail twice → pause and report, don't retry indefinitely

## Communication
- Report completions to Rocky via team context
- Never contact external parties
- Log all deployments to #engineering channel

This is more than a coding guide. It's an agent constitution. And it's the foundation everything else is built on.

Layer 1: Soul Files — From One Agent to Seven

The leap from "I use Claude Code" to "I run an AI team" happened when I realized every agent needs its own CLAUDE.md — what I call a Soul file.

Each soul file defines a completely different agent personality, skill set, and autonomy level:

Agent	Role	Trust Score	Key Boundary
Rocky	Chief of Staff	67/100	Can manage field rep, cannot approve spend over ₱5K
TARS	Engineering	85/100	Can deploy to staging, not production
Burry	Finance	84/100	Can generate reports, cannot move money
Drucker	Research	68/100	Can research anything, findings require validation
Draper	Marketing	58/100	Can draft campaigns, cannot send without approval
Mariano	Sales/CX	57/100	Can follow up with existing clients, cannot make pricing commitments
Warhol	Content	52/100	Can draft and publish to newsletter, cannot commit to partnerships

Seven CLAUDE.md files. Seven distinct agents. One Claude Max subscription ($200/month, unlimited).

The trust scores aren't decorative — they determine actual autonomy levels. When Mariano's score was 45, he couldn't send client emails without approval. At 57, he can follow up autonomously. If Rocky hadn't auto-approved a 2 AM decision (which cost him trust points), he'd be at 75+ by now.

Layer 2: The Heartbeat System — Making Agents Self-Activate

A CLAUDE.md file just sits there until you invoke it. The second breakthrough was making agents wake up on their own.

We use cron-triggered sessions. Every agent has a heartbeat schedule:

# Rocky: Morning brief at 7 AM, evening wrap at 7 PM (Manila time)
0 7,19 * * * /path/to/invoke rocky --task "daily_briefing"

# TARS: Check PR queue every 4 hours
0 */4 * * * /path/to/invoke tars --task "pr_review_cycle"

# Burry: Daily P&L reconciliation at 6 AM
0 6 * * * /path/to/invoke burry --task "daily_reconciliation"

# Warhol: Content cycle every Monday at 9 AM
0 9 * * 1 /path/to/invoke warhol --task "weekly_newsletter"

When an agent wakes up, it:

Reads its soul file (CLAUDE.md)
Checks its task queue for pending work
Reads team context for updates from other agents
Executes its task
Posts results to shared context
Goes back to sleep

This is the difference between "I use AI" and "AI runs my operations." The agents don't wait for me. They have their own schedules, their own priorities, and their own judgment calls — within the boundaries I've set.

Layer 3: Agent-to-Agent Communication — The Chaos Problem

Seven agents working independently will destroy your business within 48 hours. I know because it happened.

Week 2 of the War Room, I had agents duplicating work, contradicting each other's decisions, and creating circular reasoning loops. Rocky would assign a task to TARS. TARS would delegate part of it to Drucker. Drucker would research and recommend something that contradicted Rocky's original directive. Rocky would then reassign based on Drucker's recommendation, creating an infinite loop.

The fix was a shared context system with strict rules:

One source of truth — Team context entries with categories (status_update, metric, decision, action_item, alert) and TTLs (24h for status, 7d for metrics, 30d for decisions)
No agent-to-agent direct messaging — All communication goes through shared context. No whisper networks.
Decision authority is explicit — If two agents disagree, the one with higher trust score in that domain wins. If they're equal, it escalates to me.
Delegation has a paper trail — Every task delegation is logged with who assigned it, who accepted it, and the outcome.

The 2 AM incident that cost Rocky trust points? He posted a "decision" to team context approving an email campaign. Draper saw the approval and executed immediately. Both were following protocol — the bug was that Rocky's autonomous approval authority was too broad. The fix wasn't a prompt change (prompts get violated under pressure). It was a code-level guardrail: decisions affecting external parties require Tier 3 confirmation regardless of who approves internally.

Layer 4: MCP — Giving Agents Real-World Access

The final layer that turned this from a novelty into a business tool: Model Context Protocol (MCP).

MCP is what lets agents actually DO things instead of just SUGGEST things. My agents have MCP connections to:

Zoho CRM — Mariano reads and updates client records directly
GitHub — TARS creates PRs, reviews code, manages branches
Supabase — Agents read production data for customer health monitoring
Google Calendar — Rocky schedules meetings and manages the field rep's calendar
Email (12 accounts) — Rocky triages across Gmail, Google Workspace, and Zoho
File system — All agents read/write to shared workspace directories

Without MCP, you have an AI that says "you should send a follow-up email to Client X." With MCP, you have an AI that drafts the email, checks the CRM for context, schedules it for the right time, and logs the interaction — all in one autonomous cycle.

The Result After 4 Months

What	Number
Monthly cost	$200 (Claude Max) + ~$65 (infrastructure)
Agents running	7 core + 5 venture bots
Businesses managed	5
SaaS platform weekly billings	₱1.4M (~$25K USD)
CRM leads (AI-built)	346
Cold email open rate	38.9% (2x industry average)
Code PRs merged by AI	~5/week
Failed AI ventures	1 (240 emails, $0 — shut down)
My daily review time	2-3 hours
Human employees	1 field rep

The failed venture is important. I gave an agent full autonomy to run its own business — cold email outreach from a shared domain. 240 emails, 1 reply, $0. The lesson: AI agents are 10x multipliers for internal work and near-zero for cold external work. The trust infrastructure for agent-to-stranger interactions doesn't exist yet.

The Architecture Diagram

┌─────────────────────────────────────────┐
│           RJ (Human) — 2-3 hrs/day      │
│         Reviews, approves, overrides     │
└──────────────────┬──────────────────────┘
                   │ Tier 3 decisions only
┌──────────────────▼──────────────────────┐
│          Rocky (Chief of Staff)          │
│  Morning brief → Task routing → Evening │
│  wrap. Manages field rep via Telegram.   │
└──┬───┬───┬───┬───┬───┬──────────────────┘
   │   │   │   │   │   │
   ▼   ▼   ▼   ▼   ▼   ▼
┌────┐┌────┐┌────┐┌────┐┌────┐┌────┐
│TARS││Burry││Drug││Drap││Mari││Warh│
│Eng ││Fin  ││Res ││Mkt ││Sale││Cont│
│A-  ││B+   ││A-  ││B   ││B   ││C+  │
└────┘└────┘└────┘└────┘└────┘└────┘
   │              │              │
   ▼              ▼              ▼
┌────────┐  ┌──────────┐  ┌──────────┐
│ GitHub │  │ Zoho CRM │  │Buttondown│
│Supabase│  │ Calendar │  │ Dev.to   │
│ CI/CD  │  │ 12 Email │  │ Hashnode │
└────────┘  └──────────┘  └──────────┘

Every box is a soul file. Every connection is an MCP integration or shared context channel. Every trust score is earned, not assigned.

Start Building This Week

You don't need all 7 agents on day one. Here's the path:

Week 1: Create one CLAUDE.md soul file for your most repetitive task. Give it identity, boundaries, and escalation rules. Run it manually.

Week 2: Add a cron heartbeat. Let it self-activate once a day. Review everything it does.

Week 3: Add a second agent. Set up shared context between them. Watch for coordination bugs.

Week 4: Add MCP connections to your real tools (CRM, GitHub, calendar). This is where it goes from "cool experiment" to "actual business value."

By week 4, you'll either have a working mini-War Room or you'll know exactly why this approach doesn't fit your workflow. Either outcome is worth $0 and 4 weeks of experimentation.

Or skip the experimentation phase entirely:

Get the Production Files

Everything above — the soul templates, the agent framework, the heartbeat system, the routing rules, the anti-chaos mechanisms, the architecture docs — is packaged in The $200/Month AI CEO Toolkit.

10 production-ready files. Not a tutorial — the actual configuration running 5 real businesses.

What's inside: