There is more to code review than (automatable) detection

A recent scholarly article titled “The End of Code Review: Coding Agents Supersede Human Inspection” argues that large‑language‑model‑based coding agents have reached a level of capability that makes traditional human code review redundant. The authors trace the practice of code inspection back to Fagan’s formalisation in 1976 and note that for fifty years developers have relied on peers to examine and comment on changes before they are merged. They claim that every explicit goal of code review—defect detection, style enforcement, knowledge transfer, and situational awareness—can now be performed by autonomous agents at lower cost and higher throughput, and that the “naïve integration” of agents writing code while humans remain mandatory reviewers is a dead‑end because it fails to provide meaningful assurance and does not scale with AI‑assisted development velocity.

The critique of this thesis centres on what the reviewer calls the “substitution myth,” emphasizing that peer review involves more than a set of discrete functions that can be mechanised. Human reviewers contribute an emergent sense of comprehensibility: when an experienced engineer encounters a diff that is confusing, the very confusion signals excessive complexity, poor abstraction, or unclear intent—signals that an LLM, which can always parse the code, cannot generate. Moreover, human reviewers exercise qualified skepticism about the necessity of a change, questioning intent, scope, and appropriateness before even assessing correctness. This includes spotting missing elements, such as an unchanged error‑handling path after an API contract alteration, a class of “absence blindness” where LLMs notoriously falter because they only evaluate what is present. Reviewers also apply calibrated attention shaped by past interactions with the author, treating a novice’s first commit to a payments module differently from a veteran’s routine refactor, a nuance the article overlooks by treating all diffs as equivalent inputs.

Beyond defect detection, the article’s reduction of knowledge transfer to mere information delivery ignores the co‑active, conversational nature of code reviews. Human reviewers and authors engage in joint cognitive work that reshapes both parties’ mental models, conveys tacit organisational knowledge, and surfaces informal agreements or recent events that are invisible to tests, documentation, or version control. This contextual awareness enables reviewers to recognise risks tied to the broader system, such as upcoming deprecations or downstream dependencies, and to exercise personal accountability that goes beyond a bureaucratic compliance checkbox. The reviewer argues that these dimensions—human incomprehension signals, skeptical assessment of necessity, calibrated attention, and shared understanding—remain essential to software quality and cannot be replicated by current coding agents, suggesting that the proposed elimination of human code review is premature and overlooks critical aspects of the practice.

⚡ Effects Interpreter