Armin Ronacher: GPT 6 Astra is impressive, but the 35-hour unattended software factory doesn't work
rseroter · x · 2026-09-10
Flask creator Armin Ronacher argues AI engineering is becoming an involution-style arms race, and shares his weekend experiment running a self-directed "slop factory" on GPT 6 Astra:
- Astra is genuinely impressive at computer use, images and complex reasoning, with relentless task completion.
- But he can't figure out how to work with it for real software engineering. His factory let the model decide the whole workflow, manage its own context, keep notes in an agent-notes folder, and spawn subagents — targeting projects like "Python with virtual threads."
- His verdict: "I honestly do not need an agent to run for 35 hours on a single prompt. It clearly does not work or result in reasonable outputs." Generated code can't be merged without review.
He also references Linear's and Anthropic's automated software maintenance experiments: better than expected, but still requiring human review.
More from coding & agent
- OpenAI staff ditch Google Forms, vibe code custom forms on ChatGPT Sites — sherwinwu · 2026-09-10
- Turing launches CEO Bench: 500+ expert tasks test frontier agents in a simulated company — ecekamar · 2026-09-10
- Seroter's daily list: 5 things you shouldn't vibe code, and trusting agents with production secrets — rseroter · 2026-09-10
- Notion CEO-shared take: own your context, rent the intelligence — ivanhzhao · 2026-09-10
- Voice in the Codex app is great for math study, but bugs pile up in heavy use — SIGKITTEN · 2026-09-10
- We gave AI agents browsers, now they can just call people — Illustrious-Fig-326 · 2026-09-10