Eval Protocol: RL for agents in any language, container, or framework (github.com)

Eval Protocol (EP) is an open standard that makes reinforcement-learning fine-tuning practical for existing production agents regardless of language, container, or framework. Instead of reimplementing environments or porting services, teams wrap their agent (Python, TypeScript, Docker, remote APIs, etc.) behind a simple HTTP “rollout” interface. EP handles rollout orchestration, metadata passing, and trace storage, letting agents speak a common protocol so trainers can evaluate and optimize them without invasive rewrites. Technically, any agent that implements the EP rollout API can plug into supported trainers (Fireworks RFT, TRL, Unsloth) or a custom trainer, enabling RLFT and evaluation across multi-turn, tool-using, or distributed agents. That decoupling unlocks realistic RL experiments on production stacks, helps MLOps build reproducible, language-agnostic rollout pipelines, and lets research engineers iterate on fine-tuning without rebuilding environments. Quickstart code is available at eval-protocol/quickstart for teams ready to integrate EP into their pipelines.