How many of these could be code checks written by the agent? Have it walk the AST looking for the antipatterns - it ends up more like a custom linter, and no need to waste tokens making the poor model subjectively analyze its decisions.