A new training framework teaches AI agents to perform better inside the tools that run them, without needing to see how those tools work under the hood.
Researchers describe ClawGym II, a reinforcement learning system built for agent harnesses - the software that coordinates how an AI model calls tools and executes multi-step tasks. The setup runs tasks in isolated sandboxes for large-scale parallel rollouts, then places a proxy at the model's boundary to capture every call it makes even when the harness itself stays opaque, reorganizing those calls into tree structures to reconstruct full multi-turn interactions for training. The team adapted two established reinforcement learning methods, PPO and GRPO, to optimize over that structure, and added support for training one model across multiple different harnesses at once. Tested on the Qwen3-30A3B model, the approach lifted first-try success rates by 9.98 points through the OpenClaw harness and 14.81 points through Claude Code on the paper's own benchmark, with further gains on two harder benchmarks, JobBench and OfficeQA.
The real significance is the black-box part. Most agent training assumes control over the whole stack, model and harness together, but this framework works even when the harness is someone else's closed system. That distinction matters as more agent products, from coding assistants to office-automation tools, ship as proprietary scaffolding rather than open frameworks anyone can train through directly.
Worth noting: the gains are self-reported on the authors' own benchmark and haven't yet been tested against rival training methods or reproduced outside the lab.