The proper solution is to treat the agent generated code like assembly... IE. don't review it. Agents are the compiler for your inputs (prompts, context, etc). If you care about code quality you should have people writing it with AI help, not the other way around.