## The claim under scrutiny
A paper on arXiv, "The End of Code Review: Coding Agents Supersede Human Inspection" by Martin Monperrus, argues that large-language-model coding agents have reached a threshold of capability at which traditional human code review is no longer a necessary component of a software quality pipeline. The argument rests on two claims: every stated goal of code review can be served by agents at lower cost and higher throughput, and the naive integration in which agents write code while humans remain mandatory reviewers is a dead end because it neither provides meaningful assurance nor scales with AI-assisted throughput.
The paper decomposes peer code review into four stated functions: defect detection, style enforcement, knowledge transfer, and awareness. Because an agent can perform each one, the conclusion follows that the agent thereby has the capability to replace a human reviewer.
## What the functions miss
The counterargument is not that agents cannot detect defects or enforce style—indeed they often can. Rather, the irreducible human dimensions of review survive only in the *integration* across functions, which the decomposition erases. Several distinct capacities fall outside what automatable detection captures:
- **Confusion as signal.** When an experienced reviewer reads a diff and says "I don\u2019t understand this," the confusion is the finding. It signals that the code is too complex, the abstraction is wrong, or the intent is unclear. An LLM will always process code without the legitimate incomprehension that exposes those problems. - **Questioning necessity.** Reviewers routinely challenge whether a change is even needed: "Should this be two PRs?" or "This solves the symptom, not the problem." Such judgments precede verification and depend on a shared understanding of the system\u2019s direction, which an isolated agent lacks. - **Seeing what is not there.** Humans notice that an API contract changed but error handling did not, or that an expected branch is missing. This recognition of absence is a class of failure—absence blindness—that LLMs struggle with, because they review only what is present. - **Calibrated attention by authorship history.** A junior engineer\u2019s first commit to a payments module naturally attracts different scrutiny than a veteran\u2019s routine refactor. That calibration depends on relationship and memory of past performance, not on the diff alone. - **Bidirectionality and coactive learning.** Code review is a joint cognitive activity: the reviewer learns the author\u2019s approach and the author learns from the reviewer\u2019s questions, producing shared understanding that neither started with. An agent\u2019s summary is transmission, not conversation that changes mental models. - **Operational context beyond the repo.** Reviewers know that there was an incident in the service last Tuesday, that a downstream consumer is deprecating an interface, or that legal forbade logging a particular field. This organizational context is rarely in tests, docs, or version control, yet it shapes every decision about what to change.
## Accountability as governance, not formality
The paper treats human responsibility as a compliance artifact—a named human for legal purposes. But personal accountability shapes how a review is conducted: the "skin in the game" produces earnest evaluation rather than mechanical sign-off. An agent that approves a pull request bears no consequences and has no incentive structure that rewards a careful judgment. Redirecting ethics concerns to "requirements engineering and post-deployment monitoring" does not restore that accountability; it defers it.
## The substitution myth
Code review is not only a detection process. It is simultaneously a coordination process, a sensemaking process, and a governance process. The substitution myth follows a recognizable pattern: decompose the human contribution into measurable functions, show that a machine can replicate each function, then declare the human redundant. This pattern fails at the same point in every domain—the human contribution that mattered most was the *integration* across functions. People adapt to unplanned circumstances and contexts, and they serve the social accountability that formal gates cannot capture. Those adaptive capacities were absent from the original decomposition, and they do not reappear when machines substitute for isolated functions.
The critique therefore agrees with the paper\u2019s decomposition but rejects its inference. Agents may well serve the detection, style, knowledge-transfer, and awareness goals at lower cost and higher throughput; the unresolved question is whether the integration layer—which determines whether a change is necessary, whether it fits the system\u2019s trajectory, and whether the team retains collective ownership of its code—can be surrendered without cost.