Review Is Not a Feature. It Is the Interface to Agent Work

Watch anyone demo an agent product and the camera stays on the generation. The agent writes the code, drafts the doc, proposes the change. Applause. What almost nobody shows you is the part that actually decides whether the thing survives in an enterprise: the review surface.

I spent a session walking through the least glamorous work imaginable: making "request changes" show a red dot so a reviewer instantly knows action is needed. Letting someone deep in a file click a badge and jump straight to the relevant comment. Copying a branch name in one click. Re-requesting review to cancel out a prior request cleanly. Focus mode that changes the background so you know where you are. None of it demos well. All of it is the product.

The part most teams miss: when an agent runs in the background and produces output a human must approve, the bottleneck is never the generation. It is the review loop. If a reviewer can't see at a glance that changes were requested, can't jump to context, can't act without friction, the output piles up. Unreviewed agent work is not deployed work. It is a queue. And a queue that grows faster than a human can clear it is worse than no agent at all.

This is why I keep insisting that review is a primitive, not a screen. It is not a step you bolt on at the end. It is the interface between human judgment and machine execution, and it has to be engineered with the same care as the agent itself. The affordances that feel trivial, a badge, a jump link, a clear cancel, are what turn a plausible demo into something an enterprise team will actually run every day.

Everyone is racing to make agents generate more. The teams that win will be the ones who made the output effortless to act on. If your agent produces ten times the work but review still crawls, you have moved the bottleneck, not removed it. Build the review layer like it is the product. Because for the human on the other side, it is.