You can now train agentic AI systems end-to-end with production harnesses and real environments using standard RL tools—no need to build custom training infrastructure for each harness type.
OpenForgeRL is an open-source framework that enables training AI agents end-to-end using complex inference systems (harnesses) like Claude Code and OpenClaw. It works by running a lightweight proxy that intercepts the harness's operations while feeding them into standard RL training systems, and uses Kubernetes to scale training across remote containers.