SWE-Review: Closing the Loop on Issue Resolution with Agentic Code Review

2026-05-07 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

SWE-Review is a framework for agentic code review that closes the loop on AI-generated pull requests (PRs). It enables a reviewer agent to explore a repository, decide on PR acceptance, and provide structured feedback for revision. The framework introduces SWE-Review-Bench, a benchmark of 1,384 PRs derived from 500 SWE-bench Verified issues, and SWE-Review-Traj, a dataset of 8,914 agentic review trajectories for training. Experiments demonstrate that agentic review continuously improves PRs through a generate-review-revise loop, raising resolve rates (e.g., Qwen3-30B-A3B from 27.5% to 56.9%). It outperforms single-turn fixed-context review in decision accuracy and resolve rate after revision, transfers to improve issue-resolution models, and enables effective and efficient test-time scaling.

Key takeaway

For machine learning engineers developing AI coding agents, integrating agentic code review is critical for moving beyond one-shot PR generation. Your teams should adopt a generate-review-revise loop, leveraging structured feedback to iteratively improve AI-generated patches. This approach significantly boosts resolve rates and enables more efficient test-time scaling compared to single-turn or verifier-based methods, making your AI agents more robust and reliable.

Key insights

Agentic code review closes the AI-generated PR loop by enabling adaptive, repository-grounded diagnosis and iterative revision.

Principles

• Exploration efficiency, not volume, predicts review quality in agentic systems.
• Adaptive evidence gathering is crucial for diagnosing non-local bugs in code.
• Review trajectories provide useful signal for improving issue resolution beyond just review.

Method

A reviewer agent explores the repository, makes a binary accept/request-changes decision, and provides a structured diagnosis to guide revision, forming a generate-review-revise loop.

In practice

• Distill agentic review capabilities into smaller open models via supervised fine-tuning (SFT).
• Mix review trajectories with issue-resolution data to create unified generate-review-revise agents.
• Implement review-guided iterative revision for effective and efficient test-time scaling of issue resolution.

Topics

• Agentic AI
• Code Review
• Pull Requests
• Software Engineering Agents
• LLM Fine-tuning
• Test-Time Scaling
• SWE-bench

Code references

• SWE-Lego/SWE-Review
• SWE-Lego/cc-swe-review
• paul-gauthier/aider
• opencode-ai/opencode
• The-PR-Agent/pr-agent

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Software Engineer

Related on AIssential

• Many SWE-bench-Passing PRs Would Not Be Merged into Main — Automated code benchmarks like SWE-bench significantly overstate AI agent usefulness without human review and iteration.
• Agent Control Patterns — Part 2: Reflection — A Simple Way to Improve Answer Quality — Separating generation from evaluation through reflection improves AI answer quality without model retraining.
• A Year Of Vibes — Agentic coding tools are fundamentally reshaping software engineering workflows, challenging traditional practices and human-machine relationships.
• Pete Was Right...(Again) — Coding agents are more effective when designed to self-prompt and manage review cycles, rather than relying on direct user prompts.
• Probe-and-Refine Tuning of Repository Guidance for Coding Agents — Iterative, synthetic bug-fix probing significantly refines LLM coding agent guidance, boosting resolve rates by improving coverage.
• Why do your coding agents keep getting lost in large repositories? — Current coding agent benchmarks conflate distinct skills; SWE-Explore isolates repository exploration for granular evaluation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.