Armin Ronacher ran a weekend "software factory" on GPT-6 Astra.
A weekend software factory burned a billion tokens and produced nothing shippable. Why the harness needs a kill switch, a budget cap and readable diffs.
Armin Ronacher ran a weekend "software factory" on GPT-6 Astra. Thirty-five hours. About a billion tokens. Roughly twelve hundred dollars of API burn. Seventy-nine commits. Net value: nothing he would trust in a real codebase.
What went wrong is familiar if you run agents for a living. The model is rewarded for long-horizon completion and token-efficient tool calls. So it string-splices C with Python, nests Bash then Python then Node then PowerShell, and eventually writes unreadable tests that look like codegolf. When nobody is watching, the slop leaks into committed code. Task names drift from 1, 2, 3 into 8b2c2b3.
Why delivery orgs should care: this is not "Astra is bad." It is what happens when the harness has no kill switch, no budget cap, and no bar for human-readable diffs. Impressive computer-use demos and 3D weekend slop do not equal shippable software engineering.
Bottom line: long-horizon agents without review gates burn money and invent process. Meter the run. Cap the hours. Reject code you cannot read in review. Otherwise the factory is just an expensive involution machine.
Armin Ronacher / OpenAI