Two thirds of AI pull requests never merge

Pull requests written with AI merge at 32.7%. Human-written ones merge at roughly 84.5%. That gap comes from LinearB's analysis of 8.1 million pull requests across 4,800 teams in 42 countries, published on 4 May 2026, and it is a more useful number than anything in the productivity debate: on current evidence, two out of every three changes an agent writes never land.

The popular reading of that is "code review is the new bottleneck". The same dataset says that is wrong in a specific and more uncomfortable way.

Reviewers are not slower on AI code. They are slower to start

LinearB found AI-generated pull requests wait more than 16 hours on average before a reviewer picks them up, against roughly 200 minutes for unassisted work — more than five times longer. But once review actually begins, the AI-generated ones move through faster: about 194 minutes, compared with 252 for manual work.

Read those two figures together. The delay is not review capacity, because the reviewing itself is quicker. The delay is in front of it, in whatever makes someone pick a different pull request first. And then the pass over that code is 23% shorter than the pass over code a colleague wrote.

Our read is that this is a trust problem wearing a throughput problem's clothes. A reviewer who defers a change for sixteen hours and then goes through it faster than usual is not reviewing harder because the stakes are higher. That is the combination worth worrying about, and it is invisible in every dashboard that counts pull requests.

What shipped this month is aimed at exactly this

On 4 August 2026 GitHub published a walkthrough that opens with a 1,721-line pull request and a reviewer's reaction: "This description isn't very helpful. I'll review this later." Its fix is not a better agent. It is splitting that output into four stacked pull requests — data model, API endpoint, chat integration, UI components — each assigned to a reviewer with the matching expertise. GitHub is blunt about the failure it is addressing: the large pull request "becomes hard to review—so it just…sits there".

Five days earlier the Copilot app shipped stacked sessions on the same premise, so one agent run arrives as an ordered chain rather than a wall. And on 23 July, Linear added author attribution to documents: "Author name indicators let you check if text was written by a colleague or added by a loop", with agent edits highlighted separately.

Three features in two weeks, from two companies, and none of them make an agent more capable. All three attack the sixteen hours, not the review. That is a reasonable bet if the pickup delay is really about legibility — a reader who can see the shape of a change starts sooner.

The reading that would make this benign

There is an honest alternative, and the data does not settle it. A 32.7% merge rate could mean agents produce worse code. It could equally mean agents made pull requests cheap enough to open speculatively, so teams now throw away drafts they would previously never have written. Those two stories have opposite implications and the same statistic.

What LinearB measured is correlation between AI assistance and merge rate, not cause. Anyone telling you the gap proves AI writes bad code is going past the evidence, and so is anyone telling you it proves healthy experimentation.

What to do about it

Split the metric. Track merge rate for agent-assisted changes separately from the rest, and track pickup time separately from review time. If your pickup time for agent work is climbing while review time falls, you have the pattern above, and no amount of extra output will convert into shipped software.

The one to watch over the next six months is whether these tools start recommending what not to review. Stacking still assumes every line eventually gets human eyes, and the 194-minute figure suggests that assumption is already quietly failing. The first vendor to ship a defensible answer about which diffs can skip review will be making a claim about liability, not about tooling. We wrote about the upstream half of this shift in the workflow that kills handoff; this is the invoice.

Sources

• 8 million pull requests reveal where engineering productivity breaks down — LinearB, 4 May 2026.
• Turn one giant AI-generated pull request to a reviewable stack — Julia Muiruri, GitHub, 4 August 2026.
• Stacked sessions and pull requests in the GitHub Copilot app — GitHub, 30 July 2026.
• Text attribution and agent-assisted editing — Linear changelog, 23 July 2026.