A Major Video Site's AI Now Fixes Its Own Hidden Bugs

Imagine ordering a package and getting a tracking number. Now imagine that halfway across the country, the tracking number quietly vanishes, but the package still arrives. Nothing looks wrong — until you need to know where it went. That is exactly what was happening inside a major video platform.

The platform runs on more than a thousand small software pieces — called microservices — that talk to each other to load your video, log you in, or recommend what to watch next. Each request carries a little tag that lets engineers trace it end to end. But at certain handoffs, like custom queues and callbacks, the request kept going while the tag did not. Tests still passed, so nobody noticed. Researchers checked 1,133 services and found 673 of them had at least one of these silent breaks.

Their fix is a tool called Backstitch, an AI agent (software that can take actions on its own). It works like a detective: it replays suspicious traffic to see whether a call really belongs to the same request, digs into the source code to find the exact handoff that dropped the tag, makes a small, limited change to restore it, then replays again to prove the fix worked. Crucially, it only repairs the tracking — it does not disturb the actual work being done.

The results were striking. At 240 repaired points, broken tracking fell from about 90% to under 5%. Across 112 days of real production use, the platform's overall break rate more than halved. Why you care: broken tracking means slow bug hunts, weaker security audits, and features that fail in ways nobody can explain. This is AI quietly doing maintenance work humans struggle to keep up with.