This demo starts with a backlog of engineering tasks and follows an agent as it takes on the first one, a library migration across tens of thousands of repositories, from a single prompt.
The agent searches the code to see what the change will touch, then picks a small group of repos to start with. It works out the fix once, then turns it into a deterministic, tested recipe. From there, the process improves itself recursively. A team of agents runs the recipe, finds what breaks, writes tests, fixes the recipe and runs it again, so it gets stronger with every repo it touches. Each change comes with diffs, logs and an evidence trail, so teams can review what happened and measure the time saved.
The result is a reusable change that can run across the rest of the codebase without spending more inference, and each new task starts further ahead than the last. The same approach works for migrations, modernization and security fixes.
