The Coming Loop

The article argues that while AI coding agents are becoming increasingly autonomous through "harness loops" that keep tasks alive beyond a model's natural stopping point, this approach produces inferior code quality. The author, Armin Ronacher, contends that current models generate overly defensive, complex code with poor invariants and unnecessary abstractions, and that removing human oversight amplifies these problems rather than improving them. He emphasizes that for code he cares about, maintaining comprehension and control remains essential, and he hasn't yet found success with fully hands-off agentic systems.