Teams adopting coding assistants report writing more code and shipping at a similar rate. The work accumulated at review, and the queue became the limiting stage.
Production capacity rose and review capacity did not
Assistants shortened the time to produce a working change substantially. Nothing comparable happened to the time required for another engineer to read and understand one.
A pipeline with one accelerated stage does not run faster. Output is set by whichever stage did not improve, and here that is human review.
The queue is the visible symptom, and hiring more producers makes it worse rather than better.
Reviewing unfamiliar code takes longer
Review is faster when the reviewer can reconstruct the author's reasoning. Conventional code carries that reasoning in its structure and its rough edges.
Generated code is uniformly tidy, which removes the cues a reviewer uses to identify where the author was uncertain and where attention is warranted.
Reviewers describe having to read the whole change carefully because nothing signals which parts are risky, and that is slower than reviewing a human draft of the same size.
Change sets grew larger
When producing code was expensive, engineers naturally kept changes small. That constraint disappeared, and pull request size grew accordingly.
Review quality falls sharply with size. Past a certain point reviewers approve on the basis of the description rather than the diff, which is not review.
Teams that reimposed size limits recovered throughput, because several small reviews complete faster than one large one and catch more.
Authors verify less than they used to
An engineer who wrote code has already reasoned about the edge cases. An engineer who accepted a suggestion may not have, particularly when it looked right.
This pushes verification onto the reviewer, who is less equipped to do it, having neither the context nor the original intent.
The teams that fixed their queues generally did so by tightening what an author must confirm before submitting rather than by changing review.
The bottleneck moved rather than disappeared
The lesson organizations drew is that development throughput was never limited by typing speed, so accelerating it produced less than expected.
The scarce resources were understanding a system, agreeing on a design and taking responsibility for a change in production.
Those remain human activities, and the teams reporting genuine gains directed assistance toward the work surrounding them rather than toward producing more code.