Running the same task 18 times with GPT-6 Astra produced 18 different workflows
The author found Astra reinvented its working method each run, so moved the parts that should never change into a shared file, and the next 6 runs became stable and reproducible.
i gave GPT-6 Astra the same task 18 times and got 18 different workflows that wasn’t the part that bothered me then i noticed all six were making the same decision differently one agent researched before building another started building immediately one checked its work halfway through another waited until the end and two spent half the run collecting context they never used the outputs looked similar enough that i almost ignored it but underneath, i had six agents inventing six different ways to do the exact same job so i tried something simple i took the workflow out of the prompt the next run started with a tiny shared file containing only the parts that should never change where work enters who owns it what evidence is required when another agent takes over what has to exist before the task can finish Astra still decided how to solve the hard parts it just stopped reinventing the company every time it woke up the next 6 runs were boring and that was exactly what i wanted same handoffs same checkpoints different reasoning only where the task actually required it i used to think autonomous agents needed more freedom now i think the useful ones need to know exactly where freedom begins
This post quotes another post: x.com/lummox_eth/status/2100288151897981255