This vision paper, by authors at Ankara University, Microsoft and Bilkent University, starts from the observation that AI coding assistants speed up code production while the review pipeline that code must pass through has stayed largely unchanged. It argues that, under these conditions, code review is no longer only a productivity bottleneck but the primary control surface for the quality and accountability of AI-produced code, and that prior work improving individual review stages does not compose into effective review on its own, because the stages depend on context that crosses tool boundaries.
The paper first traces code review through five eras defined by practice: ad hoc review (1940s–1960s), formal inspection (1970s–1990s), lightweight peer review (1990s–2000s), integrated code review (2000s–2010s) and automation-assisted review (2010s–2020s). It then models the traditional pull-request workflow and groups its challenges into four areas — PR creation (missing context, missing documentation, missing issue links), reviewer assignment, code review itself (change understanding, change impact, PR-Issue Alignment, large diffs, time pressure) and comment provision (usefulness, sentiment and toxicity) — linking several of them to the LGTM Smell.
Its proposal is an AI-powered code review framework of five stages, each combining LLM agents or conventional tools with explicit human-in-the-loop quality gates, followed by a discussion of risks and implications for practitioners and researchers.
Key Points
- The paper’s three stated contributions are that AI-accelerated code production amplifies rather than mitigates the shortcomings of PR-based review, that stage-scoped improvements cannot by themselves produce effective review across the lifecycle, and a five-stage framework that operationalizes treating review effectiveness as a lifecycle outcome.
- PR Creation: agents generate the PR title and description, link or create issues, run an automated first-pass review and suggest fixes, and the author interactively revises the draft before submitting it — a mandatory human verification step.
- PR Augmentation: four analysis agents run concurrently — alignment analysis, bug-proneness analysis, runtime analysis in a sandbox, and change impact analysis — and a summary agent synthesizes their structured findings into claims paired with evidence references and confidence indicators, surfacing contradictions rather than resolving them silently.
- Reviewer Selection uses conventional reviewer recommendation techniques rather than LLMs, which the authors justify by the balance of computational efficiency and token cost.
- AI-Assisted Code Review centres on a PR Review Agent that reviewers address in natural language, and on a “diff-map” that organizes changes around logical units such as functions and classes, with each unit anchored to the analysis reports about it; the stage also includes explanation, automated review, fix suggestion, toxicity measurement and usefulness measurement agents.
- PR Retrospective generates review summaries intended for both humans and later machine retrieval, and computes review metrics for continuous process improvement.
- In the proposed workflow, final merge authority stays with human reviewers, and the authors describe reviewers as moving from manual inspectors to supervisory operators of agents.
- The discussion names risks including hallucination and context degradation, limited generalization across projects, error accumulation across agents, evaluation difficulty, transparency, accountability, privacy, Automation Bias, deterioration of knowledge sharing and implicit mentoring, and hidden economic costs from cascading agent errors, for which it suggests confidence-based circuit breakers.
Notes
The framework is a proposal rather than an implemented or evaluated system; the paper offers a research agenda instead of results. For practitioners it recommends modular, incremental adoption rather than deploying every agent at once, and treating the framework as an internal platform, and it floats dynamic risk-based review routing (“fast lanes” and “high-risk lanes”) as a possible future extension while warning that misclassification is its main danger. For researchers it calls for redefining review quality for hybrid human–AI settings, stratifying evaluations by review comment type, studying interface design for AI-assisted review, context enrichment, and evaluation metrics that capture review depth rather than speed. The paper’s title uses the term Agentic Code Review for this vision of review.