As agent benchmarks multiply, the AI research community is paying a growing integration tax, with each new benchmark requiring custom infrastructure and tooling. In this session, Alexandre Lacoste introduces Common Unified Benchmark Environments (CUBE), a universal benchmarking protocol built on MCP and Gym that defines shared APIs for tasks, benchmarks, packages, and registries. We explore CUBE’s design and show how it integrates with existing evaluation workflows, including early reference integrations validated in collaboration with NVIDIA using NeMo Gym and NeMo Evaluator. Developers will learn how to adopt CUBE, integrate it into their own tooling, and contribute to a more interoperable and scalable benchmarking ecosystem.