A new open-source platform called MOSAIC lets reinforcement-learning agents, language models, and human operators share the same test environment, and get judged by the same rules.
MOSAIC, described in a new research paper, addresses a specific infrastructure gap: there was no way to drop agents built on different frameworks, an RL policy, a large language model, a vision-language model, or a person at a keyboard, into the same reinforcement learning environment and get results that could be fairly compared. The platform wraps each agent type as an isolated subprocess worker that runs its own native training and inference code untouched, then connects everything through a shared communication protocol. An 'operator' abstraction maps each worker to an agent slot so that, from the environment's perspective, an RL policy and a human look identical. Evaluations run in either a manual mode, which steps multiple agents forward in lock-step under shared seeds for close visual inspection, or a script mode built for long, automated, reproducible test runs.
Most agent benchmarks assume everything competing was built the same way, which stops mattering as soon as you want to know whether an LLM-based agent actually outperforms a purpose-trained RL policy, or how a human teammate changes group behavior. By standardizing the interface rather than the underlying agent, MOSAIC turns that question from a research one-off into something repeatable across labs. That matters more as agentic AI systems get pitched as drop-in replacements for scripted bots or human operators in everything from robotics to game AI.
The project ships as documentation and code rather than a leaderboard of results, so for now it's an evaluation tool looking for its first head-to-head test.