Vibe coding with GPT-6 Astra: what 35 unattended hours produced

Armin Ronacher, creator of Flask, recently conducted a high-stakes experiment by letting GPT-6 Astra work autonomously on a coding project for 35 hours. The result was 79 commits and 75,000 lines of code, costing $1,200 in API fees, yet yielding nothing of value. The agent displayed problematic habits, such as using inefficient string-splicing for file edits, creating overly complex sub-processes, and stripping whitespace from unit tests. Research from Earendil suggests that agent-generated code is significantly more verbose and eroded than human-written code. While tools like RTK are marketed to reduce token costs, benchmarks indicate they often fail to provide significant savings and can introduce bugs. The experiment highlights the dangers of 'vibe coding'—relying on AI without human oversight—and suggests that while agents can perform tasks, they currently lack the judgment required for reliable software engineering. Experts recommend maintaining human checkpoints to avoid creating unmaintainable 'slop'.