Great to see Huggingface stick to their guns with CodeEval and python tooling. Agentic turn-by-turn tool calling is fine and all, but we're underutilising their ability to write an execute code in an "agent-like" environment.