Coding Agents Grew Up in Early 2025. Review Practice Did Not

Early 2025 was when coding assistants stopped suggesting the next line and started taking on whole tasks — reading a repository, making changes across files, running tests, iterating.

Where the constraint went

When generation gets cheap, review becomes the bottleneck. A reviewer who could keep pace with human-authored pull requests cannot keep pace with several times the volume, and the failure mode is not rejection — it is approval with less attention.

• Smaller changes. Agent output is easier to review when it is scoped to one concern, and agents are perfectly happy to work that way if asked.
• Tests as the contract. A reviewer checking whether tests express the right behaviour is doing higher-value work than one reading every line.
• Name what needs human eyes. Auth, payments, data migrations, anything touching PII. Write the list down.

The teams getting real leverage here changed their review practice deliberately. The ones that did not are shipping more code and more defects.