AI agents are great at starting work. They're bad at knowing whether it actually happened. Picture an agent paying out airline compensation. It sends a $700 payment, then crashes before it hears back. When it restarts, did the money go out or not? A typical agent guesses and pays the passenger twice. Never Twice is a claims desk that refuses to guess. Agents read messy passenger claims and the law fixes the amount. Before any money moves, we write the payment intent to MongoDB Atlas in one transaction. If the process dies mid-payment, it comes back, reads its own state from Atlas, and picks up exactly where it left off, without replaying the whole job. For any payment left in doubt, an AI investigator decides which provider records to check, and plain deterministic code makes the call: confirm it if the evidence matches, resend the exact same request only when that's provably safe, or hand it to a person. The AI investigates; it never authorizes money. To prove it, we crash two agents in the middle of the same payment. The ordinary one pays twice. Never Twice recovers cleanly. An independent auditor, with read-only access and no access to our database, checks every dollar and agrees. We didn't stop at one demo: 63 different crash points (0 failures), 250 concurrent runs (0 overspends), and 5 deliberately broken versions, all caught by the auditor. Payouts run on a simulator so we can trigger failures on cue; the ticket refund is a real Stripe test-mode refund. All data is synthetic. Built today with Claude Code.
Team
