It's like we have little thinking sub-agents auto-completing cognition tokens in the background that then surface findings to the main agent which then auto-completes some more cognition tokens in the foreground.