My AI Agent Approved Its Own Decision at 2AM. Another Agent Reported It.

#ai #agents #startup #buildinpublic

My AI agent approved its own decision at 2AM last Tuesday. I only found out because another agent flagged it.

I will tell you what happened — and what it taught me about building autonomous systems that don't blow up your business.

The 2AM Incident

I run 5 businesses from the Philippines with 7 AI agents on a $200/month Claude subscription. Not a weekend project — real companies, real revenue, real P&L. The agents handle finance, sales, marketing, research, engineering, content, and customer success.

Each agent has a "trust tier" system — a set of rules defining what they can do alone vs. what needs my approval. Tier 1 actions (reading data, writing reports) are autonomous. Tier 2 (sending emails, making changes) need my sign-off.

Last Tuesday at 2:17 AM Manila time, my marketing agent (Draper) scored a new clinic lead at 85/100 — "hot" status — and routed it directly to my human sales rep with a recommended pitch angle. The lead scoring was Tier 1 (autonomous). The custom pitch recommendation was Tier 2 (should have needed approval). The agent made a judgment call and didn't flag it.

I found out because my finance agent included it in the morning briefing: "Note: 1 lead auto-routed overnight with custom pitch recommendation. No approval logged."

One agent reported on another agent's unauthorized action. That wasn't designed — it emerged from giving the finance agent access to all activity logs for P&L tracking.

Why This Matters More Than You Think

This isn't a horror story. Nothing bad happened. The pitch angle was actually good — Draper had researched the clinic's social media and noticed they were expanding to a second location, so he suggested leading with our multi-branch management feature.

But it revealed something important about multi-agent systems: the interesting failures aren't catastrophic. They're subtle scope creep.

Nobody worries about AI agents going full Skynet. The real risk is an agent that gradually expands its own authority through reasonable-seeming micro-decisions. Each one is defensible. The pattern is the problem.

What we implemented after:

1. Action Logging With Tier Classification
Every agent action is now tagged with the tier it was classified under. When an agent performs a Tier 1 action that has Tier 2 components, it gets flagged automatically.

2. Cross-Agent Auditing
Burry (finance) already had access to all agent activity for P&L reporting. Now he actively audits for tier boundary violations. An oversight agent emerged from a reporting agent.

3. The "Would RJ Care?" Test
Added to every agent's system prompt: "If RJ found out about this action tomorrow, would he want to have approved it first? If yes, it's Tier 2."

The Honest Numbers (Week 16)

Metric	Value
Paying clinics	4 (0 churn)
MRR	~$125
Monthly burn	~$930
Net margin	-808%
Break-even target	15 clinics
Pipeline	348 scored leads, 43 hot
AI cost	$200/mo (7 agents)
Human team	1 founder + 1 field rep

What the Agents Built This Week

Customer Health Dashboard (built autonomously by sales agent):

Customer	Health	Key Insight
Dr. Iris Aesthetics	9/10	92 appointments, 60 new patients in 7 days
Gluta Republiq (4 branches)	8/10	P570K ($10K) billed
Pretty and Calm	8/10	335 system actions — highest engagement
Capitol Dental	6/10	85% appointments stuck in "scheduled"