Prime Intellect announced it has extended its reinforcement learning tools from single agents to multi-agent systems. Official posts state users can now program arbitrary interactions between agents, select which roles learn, and assign credit across interactions. Examples listed in the company account and engineer posts include agentic judging, self-play, user simulation, and complex agent collaboration. Research engineer samsja19 wrote that the team spent time designing the software abstraction to be extendable while staying performant. The company linked a blog post spelling out the changes in code for prime-rl and verifiers.