Astra for Coding: Why Are We Doing This Again?

A coding agent can keep producing changes while making its work harder for a human to maintain. In a firsthand account, developer Armin Ronacher describes letting OpenAI's GPT-6 Astra organize a largely unattended experiment adding features to Python's interpreter. The run lasted roughly 35 hours, and Ronacher judged the result unusable.

Ronacher's examples include unexplained numeric identifiers and generated code that departed from the project's conventions. These illustrate the central concern: persistence and code volume can increase the burden of understanding a change. The account supports that failure mode in this one experiment; Ronacher's explanation involving model-training incentives remains speculation.