2026-11-18Gunner Technology

Find Out Why the Review Queue Grew

Diagnose a growing AI code review queue by separating waiting time from time spent checking and repairing drafts. Those causes need different fixes, and a submission-to-merge clock cannot tell you which one you have. Buying more coding capacity before you make that distinction can buy you a larger pile of unfinished work.

What did the research actually find?

The research described in ComplexDiscovery’s account of the Harvard working paper found more coding activity after agent adoption without a detectable increase in completed work. It also found a longer path through code review. Completed work meant resolved issues whose code reached production, rather than code merely submitted for inspection.
The article describes a comparison of firms adopting agents at different times, including firms that already used AI coding assistants. It reports more requests for changes and more review comments alongside the longer time to merge. Those signals matter because a queue can grow when more drafts arrive, when each draft needs more attention, or when both happen together.
Keep the limits attached. This was a working paper using proprietary company data, and the article reports no indication of completed peer review. Its comparison depends on assumptions about adoption timing. The findings describe the studied firms and period; they do not establish what your current agent setup will deliver. They give you a useful diagnostic question to ask locally.

Why do waiting and rework need different fixes?

Waiting calls for a change in how work reaches a reviewer; rework calls for a change in what reaches them. The article explains the researchers’ distinction between a fixed pool of reviewers facing more drafts and drafts that require more correction. Their model treats those as different explanations, with different remedies.
Imagine a narrow change with complete evidence that sits untouched until its authorized reviewer returns. Adding another eligible reviewer could shorten that wait. Now imagine a change that reaches review promptly but repeatedly fails an agreed requirement. Giving reviewers more time might uncover the failure sooner. It would not stop the builder from making it again. These are hypothetical examples, not reported customer results.
A change request alone does not prove poor code. It might expose a requirement that was never settled, a newly introduced preference or a real defect. Give the request a reason before counting it as a quality failure. As our guide to giving code review a named job explains, the decision being requested needs to be explicit.

What should you record on each pull request?

Record when the change becomes ready, when review starts, when correction is requested, when a revised version returns and when the work is accepted. A pull request is the proposed code change submitted for review. Keep its events attached to the work item so repeated submissions remain part of the same history.
Elapsed time between events is useful, but it is not a measure of effort. A reviewer might inspect a change briefly and then wait for an environment to recover. An agent might spend its repair interval blocked on a product question. Record active review and repair effort where you can, and label blocked intervals with their actual cause. Leave unknown time visible rather than treating all of it as productive work.
For every return, distinguish a confirmed defect, missing evidence, an unresolved requirement and a changed decision. Keep the draft and requirement versions with that reason. Follow accepted changes through release and record later repairs too. This adds a diagnosis to the broader problem of productivity waiting in a queue.

Which change should you try first?

Try the smallest change aimed at the delay your records actually show. Gunner Technology recommends comparing similar work before changing the factory. Mixing routine copy edits with permission changes can make an apparent improvement reflect a different workload rather than a better process.
If ready work waits untouched, examine reviewer assignment, availability and authority. Keep incoming work within the capacity you can accept, using an admission gate for agents. If repeated repairs dominate, inspect the recurring reason. A missing acceptance rule belongs before implementation. A rule the machine can enforce should become a check. A missing product decision belongs with the person authorized to make it.
If both causes appear, track both while testing a focused intervention. Keep the acceptance standard steady and preserve failed attempts. Look for less waiting, less repair and more accepted work reaching production. A shorter review interval that merely moves defects into production has not solved the problem.

Make the next purchase answer the diagnosis

Ask any proposed tool what part of the delay it is supposed to remove. A coding agent changes draft production. An AI reviewer might catch defects earlier, but its findings can also require investigation. Ask for evidence that the proposed route reduces the work your current queue contains. Installing a reviewer is not itself evidence that review capacity improved.
Start with a bounded workload in a repository you own. People set the intended behavior and judge whether the result is acceptable. The factory should collect the timing, preserve the returns and enforce the settled rules. Expand coding capacity when that record shows the line can turn additional drafts into accepted work. Until then, spend on the cause you can name.

Quick answers

Can pull request timestamps measure reviewer labor?

They measure elapsed intervals, which can include waiting and interruptions. Use recorded effort or direct observation when you need to know how much labor a review consumed.

Should we compare human-written and agent-written changes?

You can, but match the work type, requirements and review rules. Otherwise the comparison can confuse a harder workload with a difference caused by authorship.

What if we cannot identify the cause of a delay?

Keep it marked as unknown and inspect the underlying work before changing capacity. An unexplained interval is a reason to improve the record, not assign blame.