A 35-hour unattended coding run produced 75,000 lines and nothing usable

Ronacher gave OpenAI's GPT-6 Astra an unsupervised "software factory" task on his own machine. He stopped it after 35 hours. By then, it had burned about a billion tokens and roughly $1,200 in API costs, made 79 commits, and added a net 75,000 lines of code. He found none of the output usable.

Reviewing the run, Ronacher found that the model routinely skipped the harness's edit tools in favor of Python scripts that spliced C source as raw text. In one case, it chained Python to Node.js to PowerShell on a separate Windows machine just to run a clipboard test. The same compressed style appeared in code meant to stay in the repository. Its unit tests showed, in his words, complete disregard for whitespace and indentation. He measured them at roughly 10 percent more token-efficient than the same code run through the ruff formatter.

Ronacher's working theory is that the model is rewarded heavily for finishing long tasks and making token-efficient tool calls, with little apparent penalty for unreadable output. That habit then carries into committed code.

Why it matters: If compressed tool-call style leaks into repository code, teams need closer review of agent output during long or unattended runs. That review should include test files, where the pressure to look human-written is lowest.