By September 2026, more than four in ten pull requests on GitHub started as an agent’s first draft. In March it was 17%. That’s GitHub’s own count, published last week in its Octoverse 2026 report, and it raises a plain question for anyone leading AI-enabled teams: who’s doing the code review on all of that, and what are they checking for?

What GitHub reported

The figures come from GitHub’s own platform data:

  • More than 40% of pull requests started with an agent’s first draft by September 2026, up from 17% in March. That’s about 110 million in September alone, roughly six times March’s volume.
  • 1.3 billion merged pull requests in the year to September 2026, more than three times the 402 million GitHub counted for 2024.
  • By September, 90% of agent pull requests merged within two weeks, against about 86% for the comparison group.

Two cautions. GitHub counts only the agent use it can detect, so the real share is probably higher: its own glossary calls the measure “a floor”. And the comparisons between agent and other pull requests are descriptive. A higher merge rate tells you the work got merged, not that it was right.

Why this matters for code review

Picture a kitchen that hires three more chefs but keeps one person checking every plate at the pass. The checker’s job was never “does it look good?” It’s “is this what was ordered, and is it cooked through?” With more plates arriving, two things can go wrong. The checker becomes the queue everyone waits on, or the checker starts glancing rather than checking, because nobody wants to be the one holding up service.

Code review is the pass. When an agent drafts the code, review is often the first point where a person engages with the change at all. And AI-generated code can look tidy: consistent names, neat comments, tests that pass. Appearance is exactly what a reviewer can no longer rely on.

AI code review tools can catch some problems, and GitHub notes that more tooling is being built for review. But a tool doesn’t answer the question that matters to a leader: who is accountable when a merged change turns out to be wrong?

There’s a related signal in Microsoft’s 2026 Work Trend Index. Asked which human skills matter more as AI takes on more work, the AI users it surveyed put quality control of AI output first (50%) and critical thinking second (46%). And 86% said they treat AI output as a starting point, not a final answer. That survey covers knowledge workers in general, not only software teams, but it points the same way.

My take

Add review capacity alongside your agents, and give every merge a named owner. If agent adoption goes up and review capacity stays flat, the work either queues or gets waved through. Neither shows up in a count of how many pull requests you merged.

Then change what review is for. Reviewers need to challenge the intent and the assumptions behind a change, and ask for evidence that it works, even when the code looks convincing. Three questions do most of the work:

  1. What was it meant to do? The intent, in a sentence, from whoever asked for the change.
  2. What does it assume? About the data, the users and the rest of the system.
  3. What’s the evidence it works? A test that would have failed without the change, a run against real cases, or a check by someone who knows that part of the system.

This is the same pressure as the review queue in episode 1 of the video series: when coding speeds up and review doesn’t, the queue grows. Four in ten pull requests starting with an agent is that pressure measured across a whole platform.

How you’ll know it’s working

Imagine a team that tries this for a month (an illustrative example, not a real one). You’d expect every merged pull request to have a named person who would answer for it, reviewers asking the three questions in their comments, and fewer “how did this get through?” moments after release. If the time work waits for review climbs instead, that’s the signal to add review capacity, not to review faster.

For a wider view of where AI is straining delivery across your teams, the free AI Delivery Health Check asks about the ground rules for AI, how closely work is checked before it ships, and who owns the problem when AI-assisted work breaks. What the Health Check covers explains the rest.