Most teams that adopted AI coding agents expected one thing: ship faster. A new Harvard working paper on 718 companies suggests they got the opposite of the metric that matters. Code volume rose about 30%, but delivered output barely moved — and time spent on code review jumped 49%.
The bottleneck didn't disappear. It moved downstream.
Why review became the choke point
Three forces collide here. First, generation is cheap and reviewing is not. An agent can produce in seconds what a human needs twenty minutes to verify. Second, agent-written code is often plausible — it reads well, passes a glance, and hides subtle issues in the exact places reviewers skim. Third, review capacity is fixed. You can't autoscale the senior engineer who understands the domain.
Anyone who's run a team through an agentic workflow has felt this: the queue of "ready for review" grows, PRs get bigger, and the careful review that used to happen before lunch now happens never.
What actually helps (beyond "write better prompts")
A few patterns are emerging from teams that are handling this well:
1. Make generation review-friendly. Ask the agent to keep diffs small and self-contained. A 40-line PR with a clear summary gets reviewed; a 900-line "feat: implement everything" PR gets rubber-stamped. The review burden is set at generation time, not review time.
2. Add a self-review pass before the human one. Have the agent review its own diff first — against the actual changed lines, not its memory of what it meant to write. Cheap, catches the embarrassing half of issues, and shortens the human pass that follows.
3. Require agent-written PR descriptions. Context is what makes review fast. "Why this approach" in three bullets beats any amount of code comments.
4. Measure review time, not just output. If your dashboard tracks lines generated but not time-to-merge, you're optimizing the part that's already fast.
The uncomfortable part
The paper's core finding is a management problem, not a tooling problem. Organizations bought generation speed and discovered their constraint was always verification — the one step that can't be delegated to the thing being verified. Until review gets the same engineering attention that generation got, that 30% more code is mostly inventory.
We dug into the paper's numbers, methodology, and the honest caveats in a full write-up (available in English and Chinese): https://skillstackhq.com/en/blog/ai-coding-agents-code-review-bottleneck
Top comments (0)