Code Review Becomes the Bottleneck

When agents can generate forty-five pull requests in a single day, the bottleneck is no longer writing code. It is reviewing it.

Teams running background agents against their ticket queues are discovering the output is good enough to merit review. At the companies furthest along, a majority of merged pull requests already originate in agent sessions. The PRs are not hallucinated garbage. They are reasonable code that might have made a different architectural choice than you would have. Every one of them needs eyes, and the volume compounds: the faster you merge, the more arrive.

Most AI review tools comment and suggest, but never say yes or no; the merge decision lands on a human. Forward-thinking teams are building AI-assisted review on top of AI-assisted generation: one agent writes the PR, a second plays the critic, brainstorming failure modes past the happy path and spinning up sub-agents to break each one. No human should even see the PR until it has been beaten up, even if that verification pass costs more tokens than generation did, because the scarce resource is now senior reviewer attention. The human still owns the merge, and not only for quality control. If agents merged a hundred flawless features overnight, the team would wake up not knowing what its own product does. Review is how humans keep a working model of the software they have to support, so the job shifts from reading every line to validating intent. When the diff stays connected to its session, a review comment becomes the next prompt, not a note the author has to translate.

The tooling has to follow the bottleneck, because a diff on a screen tells you what changed, not whether the feature works. The lazy version of AI review is already common: paste the diff into an agent, the agent says the code is good, approve without ever running it. The fix is to make review a shareable act: give every change a live environment restored from a snapshot, app running, code loaded, and let the agent record a video of the feature working, so proof of behavior travels with the diff. A reviewer can watch a twelve second clip in chat and approve the merge without opening the repository. Watching the feature survive its edge cases is more meaningful than reading the code that produced it.

I've made the broader case elsewhere that event-driven agents are the underappreciated capability of this wave, and review is one of the highest-leverage events to wire. A new PR is an event. An agent should already be summarizing the diff, flagging risk, and surfacing the architectural decision before the human opens the tab.

Organizations that treat AI adoption as a code generation problem will plateau at the review bottleneck. The ones that build review infrastructure alongside generation infrastructure will compound. Pick your wall now.