How to train multi-agent LLM teams end-to-end?

Most multi-agent LLM systems rely on hand-crafted prompts and rules rather than learning. UnityMAS-O treats the whole workflow—agents, tools, interactions—as a single trainable unit, letting you define agent roles, reward structures, and parameter sharing without rebuilding optimization code. On question-answering and code generation, RL training improves smaller models substantially over manually specified baselines, suggesting this approach could become a standard way to develop multi-agent systems.