slime
clean separation between optimizer loop and rollout logic feels like the scalable way to support both math RL and agent RL without framework sprawl.
Gist: the design claim is “one stable RL kernel, task-specific variety in data generation.” Training stays fixed; multi-turn tools, environment feedback, verifier rewards, and other agent behaviors are modeled as rollout/data-gen differences rather than separate trainer forks.
Newsletter angle: “Agent RL stacks may converge on a small trusted core plus flexible data-generation layers.”