Boris Cherny explains how Claude agents now run in persistent loops, pushing code beyond a single completion. The model calls tools, edits files, runs tests, and the loop continues until a harness decides the job is finished.

Cherny notes that everyday agents already embed a loop, but a second, harness‑level loop lets developers inject new messages, modify context, or hand tasks to a different machine. This extra layer has become a hot topic on Twitter and among engineering circles.

He argues that hands‑off loops produce code that is defensive, complex, and hard to understand. Modern models add local fallbacks instead of enforcing invariants, leading to bloated, fragile programs that fail to meet the quality Cherny demands.

Despite these pitfalls, loops shine in areas like automatic porting, performance testing, and security scanning where the output need not survive long. Cherny concludes that while loops help automate tedious experiments, they remain unsuitable for crafting maintainable, long‑lived code.