The Failure of Modern Code Review

Code review has become a bottleneck in software development. With code churn increasing by 861% and review times rising 4x, the process is no longer effective. Currently, over 30% of changes are merged without any review, and when AI agents perform the reviews, humans often skim the output without true engagement. This creates a false sense of security where the wrong things are being configured.

Decoupling Alignment from Semantic Accuracy

Code review serves two distinct purposes: semantic accuracy (catching bugs, security, and conventions) and alignment (mentorship, architectural feedback, and knowledge sharing). While AI can handle the former, the latter is essential for team collaboration and must be preserved.

Spec-driven development is not a complete solution because it mimics the rigid 1970s waterfall model, failing to account for the iterative, non-deterministic nature of LLM-based coding. The critical decisions—the "intent"—are currently trapped in the ephemeral prompts used during the coding session and are discarded once the pull request is opened.

A New Workflow: Intent and Evidence

To modernize the process, teams should shift the review surface from the code diff to the intent and evidence:

  • Capture the Session: Treat the back-and-forth between the developer and the AI agent as the source of truth. Convert these decisions into formal acceptance criteria.
  • Build an AI Slop Registry: Mine the last 1,000 review comments to identify recurring feedback. Codify these into a registry that acts as a persistent guardrail, ensuring that common mistakes are caught automatically rather than through repetitive human intervention.
  • Automated Verification: Use the acceptance criteria to generate a test plan that a verification system runs against a live preview. Reviewers should then evaluate the evidence (e.g., screenshots, database snapshots, or test results) rather than the raw code.

This approach moves the human role from line-by-line checkers to architects who govern intent and verify outcomes. While this transition follows a J-curve of initial effort, it ultimately scales by codifying team best practices into a system that learns over time.