GPT-6 Astra splits a single task into a four-role team
Given one task, Astra split itself into planner, researcher, coder and evaluator roles that rejected and reworked each other's output, all within one session.
i gave GPT-6 Astra one task and watched it turn itself into a 4 person team i didn’t create four separate agents then Astra realized the task was too wide for one clean run and started splitting the work itself one instance planned the route another disappeared into research another started building the last one waited to attack whatever came back then something happened i wasn’t expecting they stopped behaving like four copies of the same model the researcher found something that changed the plan the planner rerouted the task the coder kept working on the parts that were still valid the evaluator rejected one output and pushed it back into the loop all of that happened inside one session by the time the branches merged again, Astra wasn’t returning the first answer it could produce it was returning something several versions of itself had already argued over that’s the part shown in this video that i find interesting we spent years trying to make one AI smart enough to do everything GPT-6 Astra is making me wonder if the better architecture is one model that knows when it shouldn’t work alone
This post quotes another post: x.com/lummox_eth/status/2100288151897981255