What Does a Reviewer Check When Anthropic Merges 30,000 Changes a Day? (Series #AI4SDLC)
At HighLoad++ I talked about where the bottleneck moves once writing code gets faster. One place it moves to is the review queue. This conversation between Anthropic’s Addy Osmani and Sonar’s Manish Kapur at AI Engineer picks up that thread. Osmani suggests changing what the reviewer starts with. Alongside the diff, the reviewer gets check results they can examine and challenge. (Here’s my State of AI4SDLC talk at HighLoad++.)
Anthropic’s scale is striking. According to the company, it was merging about 2 500 changes a day into its monorepo in May 2026, and more than 30 000 by September. To handle the volume, it moved the monorepo from Git to Scion, its own version control system. Merge counts give you a sense of the load on the infrastructure. To understand the effect on users, you also need data about delivery and defects. I covered this story from the Anthropic engineering blog in a separate Research Insights episode.
What does the reviewer get?
According to Osmani, Claude does the first review pass on the Claude Code team. It gets to try the change much as a user would. For a UI change it can open the app in a browser, go through a scenario, and attach a recording. For another change, the result of an end-to-end test or a check of a property across different inputs might be more useful. The reviewer examines the evidence and decides whether it is enough for the change. If something breaks after the merge, the engineer who took ownership of the change is responsible for fixing it.
This shows why the verifier needed the right to say no in the Loop Engineering episode with Maxim Smirnov. Suppose the agent attaches a video of an administrator successfully exporting a report. Great, one scenario passes. What happens if a user without permission presses the same button? That is a different scenario with a different expected result.
Kapur applies this logic to the whole chain. Static analysis may find a path to SQL injection, an agent-assisted review may spot a departure from what the task intended, and a dependency check may find a vulnerable package.
#AI4SDLC #CodeReview #LoopEngineering #Agents #Engineering