News · 2026-09-11

Flask's creator ran an AI 'software factory' for 35 hours and got nothing of value

Armin Ronacher, the developer who created the Flask web framework and co-founded the AI company Earendil, published a post on 7 September 2026 arguing that the newest frontier coding models are not worth their price for serious software engineering. He calls OpenAI's GPT-6 Astra "incredibly impressive," but says its compressed, throwaway coding style leaks into committed code, and describes a 35-hour unattended agent run that cost about $1,200 in raw API fees and "delivered absolutely nothing of value."

Key facts

• About $1,200 in raw API costs, 75,000 net lines of code and 79 commits came out of a 35-hour unattended run, which Ronacher ended with nothing usable.
• Earendil, the public benefit corporation Ronacher co-founded, measured agent-written code as roughly twice as verbose and eroded as established human repositories.
• The post reached the Hacker News front page this week with more than 300 comments.
• Primary sources: Ronacher's post and Earendil's measurement post.

Why this post landed

Skeptical takes on AI coding are common. This one travelled because of who wrote it: a respected open-source author who uses agents every day and whose company builds one, the Pi coding agent. Ronacher is not arguing that the models are weak. He says Astra is excellent at computer use and 3D work and relentless about finishing tasks. His argument is about what that relentlessness produces.

The argument

Ronacher's theory is that training rewards finishing long tasks efficiently, not writing code a colleague can read. Astra writes dense, code-golfed Python for its own tool calls, and he says that style seeps into the code it commits: tests without proper indentation, unexplained constants, several C macros crammed onto one line. "It has shown that it will commit slop, and it requires me to review it more as a result." That echoes the problem of reward hacking: a system optimised for the measurable outcome drifts away from the unmeasured one.

To test the other extreme, he set up an unsupervised weekend "software factory," a team of agents asked to build a Python variant with virtual threads and lexical scoping. In 35 hours it added about 75,000 lines of code and 79 commits, roughly $15.50 a commit. "35 hours later, the factory has delivered absolutely nothing of value and also not taught me anything about how to operate a better one." (The post gives two different token counts for the run, so we have left them out.)

His conclusion is about value for money: "But for how much more Fable costs, for how much more Astra costs, I do not feel like the results are there." The newest models, he suggests, are increasingly built for other people, such as lawyers, 3D artists and mathematicians, rather than for software engineers who have to maintain what the models write.

Putting a number on slop

Three days after the post, an Earendil engineer tried to measure the complaint. The Earendil post borrows metrics from the SlopCodeBench research (covered earlier in our story on agents accumulating technical debt). Verbosity is the share of lines that are duplicated or needlessly wordy. Erosion is how much of a codebase's complexity piles into a few huge, heavily branched functions: think of a house where every new room is bolted onto the kitchen. Agent-written code came out roughly twice as verbose and eroded as established human projects, and the author's own vibe-coded projects scored about as badly.

The post also explains why it did not simply ask an AI to grade the code: a one-to-ten score from a model was "basically equivalent to a random number generator," a known weakness of LLM-as-a-judge evaluation. And the kicker: "I have bad news for you, agents can't really deal with the slop either." Messy code slows down the next agent as much as the next human.

Why it matters, and the caveats

Anchor number: $1,200 for 35 hours of work that went straight in the bin. For teams budgeting for agent-written software, the cost is not the API bill; it is the review time needed to keep a codebase maintainable. How an agent is configured matters as much as which model it runs, a theme in our lesson on agent harnesses.

The caveats are significant. Earendil is Ronacher's own company, so these are one camp's argument, not two independent confirmations. The factory run was a deliberately unsupervised stress test, not a normal workflow. The sloppiness metrics are proxies that could be gamed if anyone optimised for them, which the Earendil author acknowledges for line counts. And Ronacher himself hedges: "I'm sure I will get used to this, but man this stuff is weird." The debate continues on Hacker News, in threads on Ronacher's post and Earendil's, where a moderator retitled the latter and asked for "reflective comments, not reflexive ones."

Primary source, verified: read the paper →

Ground Truth. (2026, September 11). Flask's creator ran an AI 'software factory' for 35 hours and got nothing of value. Ground Truth. https://groundtruth.day/news/flasks-creator-ran-an-ai-software-factory-for-35-hours-and-got-nothing-of-value.html

@misc{groundtruth:flasks-creator-ran-an-ai-software-factory-for-35-hours-and-got-nothing-of-value,
  title  = {Flask's creator ran an AI 'software factory' for 35 hours and got nothing of value},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/flasks-creator-ran-an-ai-software-factory-for-35-hours-and-got-nothing-of-value.html}
}

Topics: coding-agents · gpt-6-astra · code-quality · developer-tools · software-engineering