AI coding makes PR review the new bottleneck
AI agents can now write code faster, which creates more PRs to review. The slow part is not simply reading the code. The harder work is understanding the goal of the change, checking whether it matches the ticket, and deciding which review comments are worth leaving.
A custom review tool now handles the first pass. The agent checks out the PR, reads the instructions, the ticket, the full diff, and nearby tests, then organizes the review context. It shows what changed, flags risky areas, adds file-level summaries, groups related files together, and drafts comments that can be accepted, edited, or rejected.
The agent does not publish comments by itself; the er still makes the final call. This reduces the time spent figuring out the context, but still takes longer than desired.
Key points
- AI agents can make code creation faster but increase the number of PRs to review.
- The main review is understanding intent, ticket fit, and useful feedback.
- The workflow uses an agent to read the PR, instructions, ticket, diff, and nearby tests.
- The tool provides change walkthroughs, risk notes, file summaries, grouped files, and draft comments.
- The er still approves and publishes the final review.