Rebuilding CI for AI-Scale Commit Volume — Gautam Korlam, SonarSource

A coding agent can produce changes faster than the queues that test, review and merge them. Gautam Korlam examines that mismatch through his work on Uber's SubmitQueue and Gitar, now part of Sonar. A merge queue keeps the main branch healthy by testing combinations of changes before they land, but a process designed for human arrival rates can become a bottleneck when agents create many pull requests at once. Faster code generation shifts the constraint to failed builds, repeated repairs, review and approvals. Korlam argues that the engineering system around the agent must change alongside the agent itself. The talk follows those constraints through CI and the merge path. Incremental builds, caching, sharding and parallel work reduce the cost of repeated feedback; a Rust example illustrates the difference between rebuilding everything and testing the affected work. Agents also need durable context from logs and failures so a repair loop does not repeatedly rediscover the same problem. At the approval boundary, policies should match the risk of the change, separate authorship from approval and tie the decision to the exact tested commit. Exceptions and overrides need reasons and a record that survives the agent's session. Product demonstrations show how this coordination can be automated, while the reported volume and latency figures remain examples from the speaker's systems rather than general productivity guarantees. The transferable lesson is to measure the whole path from proposed change to accepted change, including recovery and accountability, before treating more generated code as more delivered value. Speaker info: - https://www.linkedin.com/in/gautamkorlam/ - https://github.com/kageiit Timestamps: 00:00 - Coding agents move the bottleneck into CI 1:13 - SubmitQueue and testing combinations before merge 2:53 - Queues built for human speed 5:24 - The software factory and repeated repair loops 7:08 - CI failures and the feedback cost 8:11 - Incremental builds, caching and parallel work 10:16 - Enforcing policy beyond passing CI 13:28 - Risk based merge and approval policies 15:21 - Tested commits, overrides and audit evidence