AI-generated pull requests wait 4.6x longer for human review, benchmark finds

The bottleneck has moved to review

A post on dev.to by Prokop Simek of the consultancy DX Heroes argues that coding agents have shifted the constraint in software delivery: code gets written in minutes, then the pull request sits in a queue waiting for a person. The evidence he points to is substantial.

LinearB's 2026 Software Engineering Benchmarks Report, covering more than 8.1 million pull requests from over 4,800 organizations, found that AI-authored PRs wait 4.6 times longer before review begins, even though they are reviewed about twice as fast once someone picks them up. For fully agentic PRs, the gap is wider still: pickup time runs 5.3 times longer than for unassisted PRs.

Faros AI's AI Productivity Paradox Report 2025, based on telemetry from over 10,000 developers across 1,255 teams, adds a related finding: teams with high AI adoption complete 21% more tasks and merge 98% more pull requests, but PR review time increases by 91%. The same report tied AI adoption to a 154% rise in average PR size. GitHub, meanwhile, said in October 2025 that its coding agent contributes to roughly 1.2 million pull requests per month.

Why agent PRs get lost

Simek's central observation is that whether an agent's PR appears in your queue depends on how the platform attributes it, and agents disagree on that. He sorts them into three buckets.

First, PRs opened under your own account. Claude Code routines attribute commits and pull requests to your GitHub user, and GitLab Duo attributes merge requests to the human who triggered the flow for compliance reasons. The failure mode is forgetting a PR because you never actually wrote it.

Second, PRs opened by a bot as drafts with you as the reviewer. Copilot requests a review when it finishes but cannot mark its PRs ready for review itself, and GitHub prevents the person who asked for the PR from approving it, so a second reviewer must be found. If your queue hides drafts, these vanish entirely.

Third, PRs not tied to any individual. The default Claude Code GitHub Action pushes a branch and leaves creating the PR to the user; custom workflows file as claude[bot] or github-actions[bot]; Cursor's cloud agents work through an app integration and open draft PRs; Devin opens PRs under its own identity unless an admin configures otherwise. No personal filter catches these.

Four saved views that cover the gaps

Because no single filter spans all three buckets, the post recommends four saved views, each answering one question. On GitHub these can live in the new pull requests dashboard, generally available since July 9, 2026; GitLab's merge request list with Reviewer, Author and Label filters does the same job.

The first is a review-requested filter, kept draft-inclusive because cloud agents request review while the PR is still a draft. The second is an authored-by-me view, which catches agents filing under your name and, since June 18, 2026, also PRs Copilot opened at your direction. The third targets agent PRs by label: cloud agents push branches with recognizable prefixes such as copilot/, claude/, codex/, cursor/ and devin/, and GitHub's labeler action can apply an agent label based on those prefixes, with local agents instructed via files like CLAUDE.md to use an agent/ prefix. The fourth is a lead view showing everything stuck org-wide: failing CI, changes requested, or no activity for several working days, with dependency-bot PRs excluded.

Habits, not just filters

Simek pairs the queues with working habits: review the last agent PR before assigning the next task, hold two fixed review windows a day, agree on a team target for time to first response rather than time to merge, limit notifications to review requests, change requests and CI failures, and keep agent tasks small so the resulting PRs fit inside a review window.

He also built Gitwall, a free MIT-licensed macOS menu bar app that surfaces these queues as widgets for GitHub and GitLab accounts. The post is partly promotion for that tool, though the queue setup itself requires nothing beyond saved searches.

Why it matters

Generation is no longer the limiting step in AI-assisted development; human review pickup is. Teams measuring agent success by output volume can miss that review latency and PR size grow at the same time, quietly eroding the gains. The remedies here are organizational, not model-related: consistent attribution, queues that actually surface agent work, an explicit first-response expectation, and smaller delegated tasks. As agents file ever more PRs, review capacity becomes the resource teams have to manage deliberately.