Definition
Plain language
MiniMax's training stack for reinforcement learning on long agent trajectories.
As stated in the literature
A modular agent-RL system separating model generation, agent harness, and training engine through standardized interfaces, supporting white-box and black-box agents and high-throughput trajectory training.
Why it matters: Modularity lets researchers mix and match models, environments, and RL algorithms, which is essential for the experiment volume agent training demands.
For example, you can swap the underlying language model behind a Forge-trained agent without touching the trainer or the agent harness, because they communicate through standardized interfaces.