As AI agents write more of the code Roblox engineers push, the hard part is no longer generating changes; it’s knowing which ones to trust. Every change passes through layers of autonomous verification, from understanding where it came from to reviewing it with the encoded judgment of our most experienced engineers, so it meets the same high bar no matter who or what wrote it. This talk shares the architectural principles behind that system, the design decisions that let it scale across thousands of PRs a day, and the lessons we learned along the way.

What attendees will learn:

• Tracing every agent-written change back to the model and harness that produced it, and catching regressions before engineers do, by reusing existing production observability infrastructure.
• Scaling expert judgment through exemplars in a multi-agent review system that stays fast and cost-efficient at volume and gets smarter with every review.
• Building a layered, risk-aware autonomous merge system that takes routine changes from prompt to prod on their own, while humans stay in the loop for everything else.