There is more to code review than (automatable) detection

This piece challenges the claim that LLM-based coding agents can fully replace human code review by showing that framing review as four discrete functions - defect detection, style enforcement, knowledge transfer, and awareness - creates a "substitution myth." It acknowledges agents can automate many measurable tasks but argues that this decomposition misses core human contributions that are not reducible to detection or explanation. Treating review primarily as verification and throughput optimization ignores emergent, relational, and contextual aspects that shape whether a change is appropriate in the first place.

Concrete examples illustrate what agents miss: a reviewer’s expressed confusion is itself a diagnostic signal about complexity, abstraction, or unclear intent; humans challenge whether a change is necessary or properly scoped; reviewers detect absent expectations (missing error handling, contract changes) that agents reviewing only present code often overlook; scrutiny is calibrated by the author’s history and risk profile; review is a bidirectional, coactive conversation that reshapes both participants’ mental models; and reviewers bring tacit operational, organizational, and accountability contexts outside the repository. The conclusion is that automation can replicate isolated detection tasks but cannot substitute for the integrative coordination, sensemaking, governance, and skin-in-the-game that make code review effective.