Agents now write more code than people can read. The teams moving fastest changed what review is for: agents check the mechanics, each change shows evidence that it works, and people spend their attention on intent and risk.
Pull requests were always about verification. Now that agents both write and review code, human review has to change too. We help engineering teams make that change.
Reading every diff line by line, mostly for knowledge transfer and style.
A verification gate. Is this a good idea, is the approach reasonable, and does it work as intended?
One human approval on every pull request, whatever the risk.
Agentic review handles logic bugs and repository conventions. People review intent, approach, and risk.
The reviewer infers correctness from the code.
The change arrives with tests, previews, and traces that show it works, so approval is easier to give.
Every change waits in the same queue.
Well-bounded classes of changes can be auto-approved, and the rest go to the reviewer who owns that risk.
Our AI engineering acceleration work starts with a readiness pass across repo context, CI, and review, then turns the biggest gaps into a plan.
For part of the work, yes. Agents reliably catch logic bugs and convention issues. People still need to judge intent and approach, because intent is often poorly written down and agents can’t reason about what they can’t see.
Only for a class of changes you can verify well, such as small fixes with strong tests. Intercom auto-approved 19% of PRs while downtime from code changes dropped 35%. Start narrow and widen as the evidence holds up.
Make each change easier to trust: tests that run fast, previews, and an agent reviewer on every PR. Then measure where human review time goes and move it to the changes that carry real risk.
See how we work with engineering teams on AI engineering acceleration.
Tell us how review works on your team today and where changes get stuck.