Post

a horrifying realization — less structure is more scalable pretraining was easy to scale bc there’s no labels, the most you can “design” is filtering for quality RL is in an unknown state where it feels promising but it’s not yet apparent how to do it fully without structure

mr. TIM

@timkellogg.me

this is from @swyx.io in announcing that he’s joining Cognition/Windsurf it’s really good. goes into detail on model labs vs agent labs, sync vs async agents, and a whole lot more www.swyx.io/cognition

swyx.ioCognition: The Devin is in the DetailsMy thesis for the future of software dev agents. This is a hastily written blogpost done on not a lot of sleep, so pardon poor pacing and structure and typos and mistakes but just getting it out there...

9/9/25, 2:18 pm1 repost10 likes

mr. TIM

@timkellogg.me

9 Sept 2025

i was not aware of the “agent lab” distinction basically, agents are probably too hard to get right for most companies, so they work with agent labs