Astra for Coding: Why Are We Doing This Again?

Why we picked it · the editor's summary

Armin Ronacher gave GPT 6 Astra one prompt, a Python with virtual threads and lexical scoping, and let it run its own workflow with notes and subagents. In about 35 hours it burned around 1B tokens, roughly $1,200 in raw API costs, for 79 commits, a net 75k lines and about 1,400 messages between agents; by his account it delivered nothing of value. The code drifted as it ran: string splicing in Python instead of the patch tool, broken test indentation, magic numbers, task names that decayed into hash-like labels. He suspects training rewards finishing long tasks and saving tokens far more than code a person can read. Before handing an agent a weekend-sized prompt, budget the time to read what comes back.