Canyouseeme

LockSmith is an AI agent that diagnoses and repairs concurrency bugs in money-moving code, paired with an independent verifier that decides whether each fix is accepted. These bugs depend on timing, so a test can pass by luck, and an AI that checks its own fix can be fooled. Muse Spark investigates a bug on PostgreSQL, writes a test that proves it, and writes a fix, but it can’t approve its own work. A verifier with no AI retests the fix on workloads the agent has never seen, checking behaviour, safety under load, speed, and the agent’s own test. Across six planted bugs, run three times each, Muse Spark had 18 of 18 fixes accepted, at about ₹7 per bug.