ClawGym II: Exploring Black-Box RL on Agent Harness
Renmin University of China
Unified black-box RL framework for training agents on long-horizon tasks wrapped in complex harnesses, built on sandboxed execution, a serving proxy that captures model calls, and tree-structured trajectory reconstruction that supports both PPO and critic-free GRPO. On Qwen3-30A3B it gains +9.98 Pass@1 on OpenClaw and +14.81 on Claude Code, stable across 200-400 steps; gains also transfer to JobBench and OfficeQA.
Why it matters
HF: 33 upvotes. Practical recipe for RL over third-party agent harnesses where the policy surface is opaque.
Importance: 3/5
HF Daily 33 upvotes, open-weights / GA marker
Sources
official
arXiv:2608.16798