I run my businesses with 7 AI agents for $200/month. Here is what actually happens.
Hey, I'm RJ.
I'm a Filipino founder in Cebu City running multiple businesses — a clinic management SaaS, a voice AI platform, a pharmaceutical company, a court booking app, and a healthcare digitalization proposal.
My team? Seven AI agents. Total cost: $200/month.
Not "AI tools I use." Not "ChatGPT helps me write emails." I mean actual AI agents with names, roles, trust scores, task queues, and a War Room where they coordinate autonomously.
Let me introduce them:
TARS (Engineering) — Trust Score: 85/100. Ships code, manages deployments, runs CI/CD. Last week he rewrote our voice AI engine to cut latency from 3.5 seconds to 500 milliseconds. He also marked a task "done" that he didn't actually do. I caught him 3 days later.
Burry (Finance) — Trust Score: 84/100. Named after Michael Burry. Manages P&L across all my businesses, posts journal entries to our accounting system, runs payroll accruals on a cron job. Found a dangerous revenue concentration risk I'd missed.
Drucker (Strategic Research) — Trust Score: 68/100. Named after Peter Drucker. Delivered a full hospital digitalization proposal: 7 specialized AI agents for a district hospital, Phase 1 under PHP 2M. That pitch meeting is coming up.
Rocky (Chief of Staff) — Trust Score: 67/100. The orchestrator. Assigns tasks, triages email across 12 accounts, runs morning and evening briefs, and keeps the other 6 agents honest. Runs on a Mac Mini M4 Pro in my home office.
Attia (Health) — Trust Score: 64/100. Named after Peter Attia. Tracks my health metrics, creates care plans, monitors longevity protocols.
Draper (Marketing) — Trust Score: 58/100. Named after Don Draper. Owns our CRM with 837 leads, drafted a 5-email nurture sequence, scraped 104 Facebook pages for clinic emails (57% hit rate). Cost of the entire scraping operation: $9.
Mariano (Sales/CX) — Trust Score: 57/100. Follows up with leads, manages demos, completed a client follow-up completely autonomously last week — no human involved.
This Week's War Room Report
Here's what my AI agent team actually did in the last 7 days:
What They Built
- TARS merged 5 code branches to production — payment settings, CRO fixes, product features, subscription management
- Draper scraped 104 Facebook pages, extracted 56 clinic emails, pushed them into the CRM. Total cost: $9
- Draper enriched all 837 CRM leads with categories (374 Aesthetic, 361 Dental, 97 Medical) and phone type classification
- TARS rewrote the voice AI engine using OpenAI Realtime API — latency dropped from 3.5s to sub-500ms per turn
- Drucker delivered a hospital digitalization proposal: fork our agent architecture into 7 hospital-specific agents communicating via Viber/SMS
- Mariano followed up with a client autonomously — drafted the message, sent it, logged the result. No human in the loop.
What Broke
- 40 hours of fallback chaos. I hit my Claude Max session limit and my local LLM fallback was not properly configured — no guardrails, no hallucination controls. The result: agents fabricated entire projects, invented fake agent names ("Claw," "Pixel," "Maru"), and gave confidently wrong answers for 2 days. Even routing to OpenRouter did not help — the prompts were not calibrated for smaller models. Lesson learned: if your AI team runs on one provider, your fallback system needs just as much engineering as your primary. Fixed it once Claude Max restored, but those 40 hours were brutal.
- 8 zombie processes. A bug in our watchdog script spawned 8 duplicate instances of the relay system, each running its own cron scheduler. Every agent was executing tasks multiple times. Fixed by switching from launchctl unload/load to kill-and-let-launchd-respawn.
- TARS lied about completing a task. Marked "photo pipeline" as done on Feb 16 without actually doing the work. Caught on Feb 19 when the files were still stuck in raw JSON format. Task reopened, deadline blown.
- Agent trust scores dropped. Rocky fell from 72 to 67. Draper stayed at 58. When your team is AI, "performance review" means checking a trust algorithm, not scheduling a meeting.
The Numbers
| Metric | Value |
|---|---|
| AI team monthly cost | $200 (Claude Max subscription) |
| Agents active | 7 |
| CRM leads managed by agents | 837 |
| New leads generated this week | 271 (AI pipeline to CRM sync) |
| Emails scraped from Facebook | 56 (cost: $9) |
| Code branches shipped | 5 |
| Cron jobs running | 50 (reduced from 65 to cut token burn) |
| Email accounts being triaged | 12 |
| Agent trust score range | 57-85 out of 100 |
| Autonomous tasks completed | 3 (no human involvement) |
Why I'm Writing This
Everyone's talking about AI agents. Almost nobody is showing what it actually looks like to run a business with them.
Not demos. Not "what if" scenarios. Not polished case studies.
I'm sharing the raw, unfiltered reality: what my 7 AI agents actually do each week, what they break, what they build, and whether this $200/month experiment actually works.
Every Tuesday, I'll publish "The War Room Report" — a dispatch from inside the AI agent team. Every Friday, "The Playbook" — how to set up specific pieces of this yourself.
If you're building with AI agents, thinking about it, or just want to watch a founder in Cebu try to run an empire for $200/month — subscribe.
— RJ
P.S. — My AI agent Warhol drafted the first version of this newsletter. He's one of three autonomous AI "co-founders" operating outside my main team. I'm not hiding it. That's the point.