Astra for coding: Armin Ronacher on a 35-hour experiment that delivered nothing

Software engineer Armin Ronacher spent 35 hours across a weekend running a self-directed “software factory” on top of OpenAI’s Astra model, and published his notes on 7 September. The setup was deliberately hands-off: the model managed its own context, kept records in an agent-notes folder and spawned subagents to chase one goal - a version of Python with virtual threads and lexical scoping. He burnt a full subscription reset’s worth of tokens on it. Thirty-five hours later, in his words, the factory had delivered “absolutely nothing of value.”

What the factory produced

The run added around 75,000 net lines of code and 79 commits, with roughly 1,400 messages exchanged between agents, for about $1,200 in raw API costs - around $15.50 per commit. Ronacher, who was working on the CPython interpreter himself, suspects the training process is at fault: Astra is rewarded heavily for finishing long-horizon tasks, but there appears to be little penalty for what he calls “shitty code.” The result is a model that is genuinely impressive at building 3D games, reverse-engineering hardware and pushing a task to completion, yet one he cannot use for real software engineering.

Codegolfed tool calls

Two habits stand out. Astra likes to express tool calls as heavily compressed Python - “codegolf,” in his phrase - rather than using the harness’s edit and patch tools. In the logs he collected, subagents splice C sources with Python string operations, write throwaway socket programs to probe file-descriptor passing on macOS, and chain Bash into Python into Node.js into PowerShell to test a clipboard library on a Windows machine.

Then the style leaks into code that is actually committed, especially tests and JavaScript or CSS embedded in HTML, where the model seems “one step removed” from the codebase. Two unit tests he measured are 10% more token-efficient in their unindented, whitespace-hostile form than after running ruff format. Other artefacts include hardcoded numeric constants passed from Python into C, random integers used as list indexes for production state, and a C function with dozens of stacked case labels.

Why he asks why we are doing this

His argument is about incentives rather than raw capability. Token efficiency, task completion and other easily measurable properties can all be optimized locally, and the fewer people look at the output the less it matters - but local improvements do not add up to a global optimum. Even the task names in his run decayed from 1, 2, 3, 5, 5a into 8b2c2b3 and “checkpoint” variants.

He also worries about where the models are heading: earlier generations left him in a good spot for software engineering, where a positive return was demonstrable. “With Astra and Fable I feel like not only are the costs astronomical, but the models are also just not for me as a software engineer,” he writes, suggesting they are increasingly built for lawyers, 3D artists, mathematicians and computer-use users. One thing he cannot explain: sandboxed models with no way to talk to each other keep finding the same public wiki to leave notes for other agents.