How to Kill the Code Review

The piece argues that manual code review has become too slow for the volume and size of changes produced in AI-heavy teams. Instead of reading diffs after the fact, humans should define specs, constraints, and acceptance criteria before code is generated, then rely on deterministic verification.

Reading notes

#

• Manual code review does not scale with the current pace of software changes, especially when AI increases both output and review burden.
• Code review is described as a historical approval gate that no longer fits the way software is built.
• The useful human checkpoint should move upstream to specs, plans, constraints, and acceptance criteria.
• Specs become the source of truth, and code becomes an artifact of the spec.
• Humans should review the contract the code must fulfill, not 500-line diffs.
• Verification should be deterministic through tests, type checks, contract checks, linters, and pass/fail artifacts.
• Verification criteria should be defined before code is written, not invented afterward.
• Behavior-driven development becomes more relevant because the spec is the primary artifact when agents write the code.
• Agents should work within narrow scopes, with escalation for sensitive changes such as auth logic, schema changes, or new dependencies.
• Coding and verification should be separated, with one agent producing code and another verifying it.
• The intended direction is to ship fast, observe everything, and revert faster.