Problem
In 2018 TSB migrated to a new core banking platform and it went badly wrong: customers locked out, payments failing for weeks, and £48.65m in fines. Every time a bank replaces its payment switch, the new one has to match years of edge cases in the old one exactly.
Approach
A tester writes the payment rules in plain English. Planner and generator agents on Vultr Serverless Inference write ISO 8583 test cases, and nothing runs until a human approves every test. Approved tests run as throwaway Kubernetes Jobs under gVisor on Vultr Kubernetes Engine, in two separate node pools, each message sent to two copies of the legacy switch and the new one. A triage agent probes for the failure boundary inside the human's bounds, and a reporter files a GitHub issue and gates the release.
Results
10,022 tests in 42 gVisor sandboxes. It caught the seeded defect: the new switch approves a duplicate $250 purchase retried 5 seconds later, which the legacy switch rejects, and triage found the boundary (retries 1 s or more apart get approved).