Agent PRs wait longest because nobody owns or checks them before review
A KaaS newsletter piece pulls together benchmark data, studies and maintainer policies on why agent pull requests stall: at the 75th percentile they sit 17.6 hours before first review, against 3.4 hours for unassisted ones. Its fix is to make the author verify, explain and shrink the change first.
Paramanantham Harrison, founder of KaaS.team and a former head of software engineering, argues that the review bottleneck isn't slow reviewers. It is pull requests that reach a reviewer with nobody having checked them. The piece follows one invented example, an agent opening a CSV-export PR overnight from a low-priority backlog ticket, and attaches sourced figures to each step.
Citing LinearB's 2026 benchmark of 8.1 million pull requests, agent PRs at the 75th percentile wait 17.6 hours before a reviewer first opens them, compared with 8.3 hours for AI-assisted PRs and 3.4 hours for unassisted ones. They are also larger, at 293 lines against 157, and merge within 30 days less often. A longitudinal study of one company (He et al., arXiv 2607.01904) put the delay elsewhere, after first review rather than before it, and found reviewers kept up by reading less: the share of PRs with a human-written review comment fell from 39% to 21% under an AI mandate. Harrison also cites a result that only 19.2% of the best AI reviewer's valid comments led to a code change, compared with 60% for human reviewers.
His fix puts the checking on the author's side. Every agent PR gets a named owner who reads every line, writes the description in their own words, and attaches proof and steps to reproduce. Tests, lint and the AI reviewer must pass before a human is asked to look, and the bot's notes go to the author. He is candid that no study he found measures whether this shortens the wait. The closest evidence is a randomized Meta experiment where routing AI-written fixes to authors rather than reviewers kept acceptance similar without adding review time.
Why it matters for code review
Overnight agents make the morning queue longer, and this piece explains why the extra PRs sit there: work nobody asked for today has nobody to vouch for it. A harness that runs all night only helps if each PR it produces arrives with an owner and evidence, so the reviewer isn't the first person checking it.