Comment by 2001zhaozhao

The directionally interesting part is that according to the announcement, K2.5 seems to be trained specifically to create sub-agents and work in an agent swarm usefully. The key part is that you don't need to manually create or prompt sub-agents, K2.5 creates them automatically, so from the looks of things it's similar to Claude Code dynamic sub-agents except the model is trained to scale to many more agents autonomously.

I wonder whether Claude is doing the same kind of training and it's coming with the next model, and that's why the agent swarm mode in Claude Code is hidden for now. We might be getting very very good agent orchestrators/swarms very soon.