ClawGym II paper: Improving agents via mixed-harness training

omarsar0 · x · 2026-08-19

The paper 'ClawGym II' introduces a method to train agents using existing code harnesses like OpenClaw and Claude Code. By placing a serving proxy at the model boundary, the system captures tool calls and organizes them into prefix trees, allowing PPO and GRPO to optimize over the multi-turn structure. Results show Qwen3-30A3B gained significant Pass@1 improvements (+9.98 via OpenClaw, +14.81 via Claude Code). The study also demonstrates that mixed-harness training leads to policies that generalize across different execution systems rather than overfitting to a single one.

Related event: RUC's ClawGym II Trains Agents via Black-Box RL on Real Harnesses(3 posts)→

Original post →

More from coding & agent

coding & agent channel →