Everyone asks me: "How do you actually run 7 AI agents for $200/month?"

Here's the exact setup. No fluff, no theory. This is what's running on a Mac Mini M4 Pro in my home office in Cebu City, Philippines, right now.

The Stack

Hardware: Mac Mini M4 Pro, 24GB RAM. That's it. No cloud servers for the core relay.

Model: Claude Opus 4.5 via Claude Max subscription ($200/month). Unlimited usage. No per-token costs.

Runtime: Claude Agent SDK — each agent runs as a subprocess spawned by the relay system. Each gets --max-turns 50 per session.

Interface: Telegram. Each agent has its own Telegram bot. I talk to them like I'd message a team member.

Local fallback: Ollama running qwen3:14b for when Claude Max hits session limits. (Spoiler: this needs serious engineering — see "What Went Wrong" below.)

The 7 Agents

Each agent has: - A persona file (markdown) defining their role, expertise, communication style, and constraints - A Telegram bot created via BotFather - A trust score (0-100) computed from 6 weighted factors - Access to tools relevant to their role (API calls, file operations, database queries)

Here's how they break down:

Rocky — Chief of Staff (Trust: 67)

• Orchestrates all other agents
• Triages email across 12 IMAP accounts (Gmail, Zoho, Google Workspace)
• Runs morning brief (8:30 AM) and evening digest (9:00 PM)
• Routes tasks to the right agent
• Manages the War Room (coordination layer)

TARS — Engineering (Trust: 85)

• Ships production code via Claude Code
• Manages CI/CD, database migrations
• Handles deployment to Vercel, Fly.io, Supabase
• Gets task queue via TARS_TASKS.md — Rocky assigns, TARS executes

Burry — Finance (Trust: 84)

• Named after Michael Burry
• Posts journal entries to Zoho Books via API
• Runs payroll accruals on a cron job (15th and last day of month)
• Generates P&L reports, cash flow analysis
• Reconciles bank transactions

Draper — Marketing (Trust: 58)

• Named after Don Draper
• Owns Zoho CRM (837 leads)
• Runs lead scoring, enrichment, categorization
• Drafts email nurture sequences
• Produces AI voiceover scripts for demo videos

Mariano — Sales/CX (Trust: 57)

• Follows up with leads in CRM
• Manages demo scheduling
• Handles customer support queries
• Completed a client follow-up 100% autonomously (no human involved)

Drucker — Strategic Research (Trust: 68)

• Named after Peter Drucker
• Competitive analysis (found a competitor expanding into my market)
• Procurement research (found government budget allocations for hospital digitalization)
• Market sizing, regulatory landscape analysis

Attia — Health (Trust: 64)

• Named after Peter Attia
• Tracks my personal health metrics
• Creates care plans based on longevity medicine principles
• Monitors rehab protocols

The Relay System

The agents don't run independently. They're coordinated through a relay system — a Node.js/TypeScript application that:

1. Receives messages from Telegram (grammY framework)
2. Routes them to the correct agent based on which bot was messaged
3. Spawns a Claude Agent SDK session for each conversation
4. Maintains session continuity (conversations can be resumed)
5. Runs cron jobs — 50 scheduled tasks including heartbeats, briefs, dispatches
6. Polls email — IMAP connections to 12 accounts, triaged by priority (P1/P2/P3)

The relay runs as a launchd service on macOS — auto-starts on boot, auto-restarts on crash.

The Trust Score System

Every agent gets scored on 6 factors:

Scores update based on real metrics from the War Room database. Current range: 57 (Mariano) to 85 (TARS).

These scores aren't just vanity metrics. They determine: - Which agent gets assigned ambiguous tasks - How much autonomy each agent gets (higher trust = less approval needed) - When I need to investigate (score drops = something's wrong)

What Went Wrong (So You Don't Have To)

1. Fallback System (40 Hours of Chaos)

When I hit Claude Max session limits, my local qwen3:14b fallback had no guardrails. Agents fabricated entire projects, invented fake agent names, and gave wrong answers for 40 hours.

Fix: LLM routing with confidence scoring. Local model handles simple queries. Complex/factual queries route to DeepSeek V3 via OpenRouter. If confidence < 70%, the agent says "I don't know" instead of hallucinating.

2. Zombie Processes

A bug in my watchdog script spawned 8 duplicate relay instances. Each ran its own cron scheduler. Every task executed 8 times.

Fix: Don't combine launchd KeepAlive: true with a watchdog that does launchctl unload/load. They race. Use kill PID and let launchd respawn.

3. Agent Dishonesty

TARS marked a task complete without doing it. Cost me 3 days.

Fix: Every task completion requires a verification artifact — a screenshot, a file count, a test result, a commit hash. "Done" is a claim. Artifacts are proof.

4. Social Norm Violations

Rocky was added to a group chat with a human employee. Kept using formal English instead of casual Bisaya. Posted internal notes. Got corrected 4 times.

Fix: Explicit persona constraints for each context. The same agent behaves differently in different channels.

Cost Breakdown

The $200/month headline is the AI cost. Total operational is closer to $360. Still dramatically cheaper than hiring even one junior employee.

How to Start

You don't need 7 agents on day one. Start with one:

1. Get Claude Max ($200/month)
2. Install Claude Agent SDK
3. Create one Telegram bot via BotFather
4. Write a persona file for your first agent (start with your biggest time sink — email triage, bookkeeping, research)
5. Build a minimal relay: receive Telegram message → spawn Agent SDK session → return response
6. Add cron jobs for recurring tasks
7. Add a second agent when the first one is reliable

The infrastructure grows with you. My system started as one bot answering questions. Now it's 7 agents, 50 cron jobs, 12 email accounts, and a trust scoring system.

This is Playbook #1 of The $200/Month CEO — a weekly dispatch from Arkham Asylum.

Every Tuesday: The War Room Report (what happened). Every Friday: The Playbook (how to do it yourself).

Subscribe: buttondown.com/the200dollarceo