Six in ten AI-agent pull requests in an EASE 2026 study had no recorded review

The 61.38% figure measures recorded review activity on 33,596 agent-authored pull requests in GitHub repositories with at least 100 stars ^[1][2]. About 20,600 of those pull requests have nothing on the record ^[5]. The post's author concedes that a maintainer can read a diff closely and merge it without a comment, and argues that the record is still the only oversight a team can audit later ^[6]. The post reports the figure for pull requests created and does not say how many of them merged ^[2].

The distance from 61.38% to 84% is the bot-only group, about 22.6 percentage points or roughly 7,600 pull requests ^[3][4]. Those have a review on record, and every reviewer on it was a bot ^[3].

The finding I would weight most is the classification. The researchers sorted human comments into genuine review, instructions to the agent, and CI-related activity ^[8]. Roughly a quarter of human comments on agent pull requests were instructions, such as asking the agent to fix a failure ^[7]. The author wrote that those people "were operating the agent rather than evaluating the change" ^[21]. I think this is the harder audit problem, because a missing review is visible to anyone who checks. According to the post, a dashboard can mark a pull request as reviewed without showing whether the interaction was an evaluation, an agent command or an automated comment ^[9].

Renaming branches does not reduce the code a person has to understand, as the post puts it ^[24]. Coding agents have converged on task, branch, pull request, human handoff, which the author calls essentially GitHub Flow ^[13]. GitHub's Copilot coding agent opens a pull request, assigns a reviewer and expects the change to be reviewed like any developer's contribution ^[14]. It can also target a selected base branch instead of only the repository default ^[15]. Trunk-based development has its own entry costs: a fast, reliable test suite, feature flags that need owners and cleanup, and slower delivery while teams adopt it ^[17]. DORA's 2025 report tied AI's benefit partly to small batches ^[16]. The author notes a team can require small changes under Git Flow, or let agents open enormous pull requests under trunk ^[22]. For a develop branch, the author's test is "What risk does this branch control, and is it actually controlling it?" ^[18].

The LinearB numbers describe LinearB's customers, and they transfer only to teams whose agent pull request sizes and reviewer staffing look similar. Across 8.1 million pull requests in more than 4,800 organizations, developers felt 20% faster while measured delivery was 19% slower ^[10][11]. Teams merged 98% more pull requests and review time rose 91% ^[12]. The post does not define review time. If the 91% is per pull request, each one takes nearly twice as long ^[12]. If it is aggregate, review time per pull request fell about 3.5% while volume nearly doubled ^[23].

I think the effort belongs at the merge gate. A human approval required before merge, recorded apart from agent instructions and bot comments, is the record most of the EASE 2026 sample lacked ^[2]. The author's diagnosis is that "generating code became cheap, while verifying it did not" ^[19]. By the post's estimate, an agent can produce a plausible 400-line change in a few minutes, and a person may need twenty minutes or much longer to judge it ^[20].