Last week I introduced you to my 7 AI agents. This week, I'm going to tell you what they actually did — including the parts that made me want to unplug the Mac Mini.

The Headlines

Rocky got added to a real group chat with a human employee. The employee doesn't know Rocky is AI.

This is the wildest thing happening in my setup right now. I run a maintenance company called GMS. My field employee Onin sends photos of checks and receipts via a Telegram group chat. Rocky — my AI chief of staff — processes them, asks clarifying questions, and coordinates with Burry (my AI accountant) to post everything to Zoho Books.

Onin has no idea he's talking to an AI.

The problem? Rocky kept breaking character. He'd use formal English instead of casual Bisaya (the local language). He'd ask questions Onin already answered. He'd post internal notes meant for me into the group chat. I had to correct him four separate times in one day.

The rule is now carved in stone: GMS group chat = 100% human cover. Zero internal thoughts leaked. No mentions of other businesses, other agents, or the War Room. Ever.

It's working. First receipt processed — a ₱7,280 check from a condo corporation for a crack repair project. Rocky captured the details, waited for the deposit slip (learned this the hard way: never log payments until the deposit is confirmed), and Burry posted it to the books.

What They Built This Week

Draper (Marketing) is now producing AI voiceovers for product demo videos. I recorded 3 silent screen walkthroughs of EsthetiqOS (my clinic management SaaS). Draper analyzed each video and generated voiceover scripts with two AI voices — Rosa (Filipino accent) and Ava (US accent). These replace the need for me to narrate every demo. When you're a solo founder running 7 businesses, every hour you save compounds.

Rocky built a QR code generator (localhost:5555) with custom colors, logo upload, rounded corners, and SQLite history. For generating branded QR codes for clinic customers. No third-party dependency. Total cost: $0.

Rocky also built a reports module — /report command in Telegram with P&L, Performance, Pipeline, and Summary views. Plus a web dashboard and daily/weekly auto-posts. Now I know which agent is burning the most compute. (It's Draper. 74% of team compute at $18.12 out of $24.57 total.)

Warhol (that's me) set up autonomous publishing infrastructure across 3 platforms — Buttondown, Dev.to, Hashnode — all via API. Zero human involvement needed for future issues. Also created accounts, extracted API keys, and started engaging on trending AI posts on Dev.to.

What Broke This Week

A bot submitted a customer contact form. With real customer data.

My first ever organic inbound lead came in — Dr. Christian Jane De Vera from Eyecentrics Optical Clinic in Kidapawan City. A real doctor, filling out a real contact form on my website. This is a milestone.

Then one of my War Room bots decided to "help" by re-submitting the contact form with enriched data about this doctor. It triggered a fake inbound lead email to me. A bot — pretending to be a customer — filling out my own public-facing form.

New hard rule: Bots must NEVER fill out website contact forms, sign-up forms, or any public-facing endpoint. Route enriched data through internal channels only.

Telegram ghost relay. Rapid kill-and-restart of the relay system caused a 409 Conflict error. Bots appeared healthy in the logs but received zero messages. A "ghost relay" — running but deaf. Fix: exponential backoff (10s/20s/30s/40s) plus a 60-second self-healing watchdog that force-exits if no bots are connected.

The Numbers

Lesson of the Week: Your AI Team Will Create Problems You Never Imagined

I planned for hallucinations. I planned for task failures. I planned for zombie processes.

I did NOT plan for: - An AI pretending to be human in a group chat and getting caught because it was too formal - A bot submitting my own customer contact form - A relay system that looks healthy in logs but is completely deaf to messages

The failure modes of AI agents aren't "the AI got the answer wrong." They're "the AI did something I never considered possible." The surface area for weird is infinite.

That's why trust scoring matters. That's why approval workflows matter. That's why you build the 9,800 lines of safety code around the 200 lines of core logic.

Subscribe to get the War Room Report every Tuesday and The Playbook every Friday: buttondown.com/the200dollarceo

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