Jev and Opus run 48 agents to catch a $49 duplicate-charge chain
Forty-eight agents walk every signup, billing and recovery path in a 340k-line subscription app, replay the cancel-reopen-checkout double charge, and stage a four-step patch proven by regression orbits.
WHOEVER LEAKED THIS JEV + OPUS SUBSCRIPTION REGRESSION DESERVES THE CFO'S THANK YOU NOTE I thought catching a $49 double-charge meant waiting for the chargeback, then eating the fee plus the refund. Apparently not. This setup is much more interesting: → JOURNEY FIELD runs 48 agents across the full 340,000-line subscription app, every signup, billing and account-recovery path walked at least once, no path skipped → REPLAY CHANNEL catches one specific chain: cancel subscription → reopen old checkout tab → payment accepted → $49 REPEAT CHARGE, with step traces and screenshots attached → PATCH REGISTER writes a 4-step fix (01 cancel subscription, 02 invalidate stale session, 03 reject old checkout, 04 retain regression test), each step scoped to its own module → TEST ORBITS compares old (fail) vs patch (pass) with a separate reviewer per orbit, no "same model grading itself", patch commits only when the old orbit still fails → RETURN SIGNAL re-runs the original user journey with replay steps and screenshots to prove the test catches the original failure, flags 7 patches that would have let the stale checkout fire a second charge so you get: Journeys walked. Bugs replayed. Patches staged. Returns proved. the interesting part isn't more test coverage. it's giving every bug a replay the patch has to still catch on commit. Instead of a chargeback plus fee plus customer-support ticket, you point 48 agents at 340k lines, catch the exact cancel-reopen-charge chain, and ship a 4-step patch that invalidates every stale checkout on day one. Replay the chain. Stage the patch. Prove the return.
This post quotes another post: “Long read: how to build a Jev-powered agent stack”