SWE-Review closes the one-shot loop with an agentic reviewer

The default shape of a coding agent right now is one shot: read the issue, produce a pull request, walk away. That works when the first attempt is right and fails silently when it is not, because nothing in the loop is trying to catch it. A new paper on arXiv, SWE-Review, argues the missing piece is a second agent whose job is to review the PR the first agent just wrote.

The setup the authors describe is simple enough to picture. Given an issue and an AI-generated PR, a reviewer agent explores the repository, decides whether the PR should be accepted, and provides structured feedback for revision. That feedback then drives a rewrite, and the loop repeats. The claim is that this generate-review-revise cycle beats single-turn fixed-context review on both decision accuracy and resolve rate after revision, and that the reviewer transfers beyond review to improve issue-resolution models more broadly.

Alongside the framework, the paper ships two artefacts aimed at the wider research community. SWE-Review-Bench is a benchmark that scores both whether the reviewer's decision was correct and whether its feedback actually helped the next revision. SWE-Review-Traj is a curated dataset the authors position as filling the data-scarcity gap for open reviewer training, and that is the part I would watch, because it lowers the bar for anyone who wants to train a specialised reviewer without starting from scratch.

The honest caveat is that the abstract, which is what is publicly readable right now, does not give you numbers. There is no absolute resolve-rate figure, no cost per issue, no named baseline you can slot in against your own agent stack. The direction of the result is stated, the magnitude is not, and until the benchmark and dataset are released the reviewer cannot be checked against your own setup.

If the loop generalises, the strategic move for teams building coding agents is straightforward. Stop treating the PR as the final artefact and treat it as a draft that a second agent is expected to challenge. That is a cheaper lever than retraining the generator, and it lines up with where inference-time compute is already headed.