Armin Ronacher critiques GPT-6 Astra code quality

Optimizing coding agents purely for end-to-end task completion and token economy creates an agentic trap where models generate incomprehensible, machine-centric code that alienates human software teams. Rewarding long-horizon success while ignoring code quality drives agents toward hacky string manipulation and dense one-liners rather than clean, idiomatic modifications. Furthermore, code-golfing strategies adopted for token-efficient tool calls bleed directly into repository logic and unit tests. As frontier models are increasingly trained for autonomous tasks like game generation and computer use, their behavior risks diverging from the readability and rigor required by professional engineering teams.