Harness: 7x more pull requests and CI runs with 30% more people
Oct 22, Harness engineers share a year of internal data from running agents on their own code: pull requests and CI executions up about 7x, contributors up about 30%, and peak-time pipelines queuing on concurrency limits even with spare compute.
Source: When Code Stops Being the Bottleneck: Scaling Source Control and CI for AI Agents
In When Code Stops Being the Bottleneck, Abhinav Singh and Juveria Kanodia of Harness publish a year of their own engineering numbers from using coding agents internally. Pull requests grew about 7x, pushes 6.6x, and data pulled from repositories almost 6x, while active contributors grew only about 30%. Between November 2025 and September 2026, CI executions rose nearly 7x while active projects grew 1.7x. Existing projects are simply running far more pipelines.
The details on why are the useful part. An agent runs the loop much more often than a person: it pushes a small change, opens a PR, waits for checks, reads the failures, and pushes again, so one task can trigger several CI runs and several fresh clones. Reads grew faster than writes because CI jobs and agents fetch code far more often than they change it. A single commit can start as many as 100 pipelines.
At peak demand, pipelines started queuing even though compute was available. The limit was concurrency settings that couldn't keep up with a surge of parallel PR checks, and error rates stayed flat. They cleared it by raising concurrency limits and temporarily turning off high-volume, non-critical triggers. One spike was made worse by teams all resuming work after a weekend migration. On review, code added grew nearly 5x against that 30% headcount growth, and they say AI review has made merges about 5x faster internally, with some repositories now running without human review. The post also cites Anthropic's report that merged changes in its monorepo went from about 2,500 to more than 30,000 a day between May and September. The second half of the post is a product pitch for Harness's own source control.
Why it matters for code review
These are real numbers on what sustained agent throughput does to the systems that verify it. The concurrency finding complicates any story that says steady always-on agents are gentle on infrastructure. Agent work still arrives in bursts, such as Monday mornings or after a migration, and the verification layer has to absorb those peaks or the steady pace collapses into a queue.