A cheap, fast model review and work tends to flag everything, including things that were done on purpose, because it can't read the context around a decision. A smart model flags fewer things and gets them right. Now, picture a cheap checker inside a full assembly line. Agents burning time and tokens fixing things that were never broken across a dozen stations at once, with no way to tell which station started the mess. The checker sets the quality of the entire line. It's the one place where saving a few cents costs you everything. Anthropik's own team stacks checkers. They chain a code review skill, a simplify skill, a verify skill and a design skill.

But the new wave of software engineering agents like Devon, Autocode, and internal enterprise agentic pipelines takes this to an entirely new level. You can assign an AI developer agent a complex GitHub issue ticket that says, fix the memory leak in the authentication microservice and update the unit tests. The agent will independently clone the repository, navigate through thousands of lines of code across multiple files, locate the source of the memory leak, refactor the code, write fresh unit tests, run the build pipeline locally to verify that no existing features broke, and submit a fully documented pull request for human peer review. What used to take a junior engineer hours or days of tedious troubleshooting can now be packaged and ready for review in minutes.