A multi-agent research system ran in production for 11 days. Two of its four agents had locked into a recursive verification loop, passing the same clarification request back and forth around the clock. Every health check passed. Bill: $47,000. Discovered when a human opened the invoice.
This is becoming a 2026 pattern. I've been reading every public agent-failure story I can find — they cluster cleanly into three failure modes, and every single one is preventable with five pieces of unsexy infrastructure most contractors skip.
🧪 In today's newsletter I broke down the 5 questions every agency owner should ask their AI contractor before signing the next build retainer. The bonus question at the end is the one that catches the bluffers.
🤔 Curious — if you've signed an AI build retainer in the last 12 months, which of these 5 questions did you actually ask, and which slipped through? What's working for you?