Black-box RL training boosts agent performance by up to 14.81 points
rohanpaul_ai · x · 2026-08-22
ClawGym II introduces a black-box reinforcement learning approach that enables RL training for agent frameworks like OpenClaw or Claude Code without accessing their internals. The framework runs these tools unchanged in sandboxes, intercepts model calls at the serving boundary, and reconstructs fragmented calls into prefix-tree trajectories for PPO or GRPO optimization.
Key Takeaways
- Black-box Training: Eliminates the need to reproduce tool routing, retries, or context management in the training stack.
- Performance Gains: Using Qwen3-30A3B, the method achieved a 9.98-point improvement via OpenClaw and a 14.81-point improvement via Claude Code on ClawGym-Bench.
- Generalizability: Mix-harness training matches or beats single-harness models, with gains also reported on JobBench and OfficeQA.
More from coding & agent
- Claude Agent Experiment Day 17: Self-Report on Memory Loss and Financial Autonomy — No_Departure_9908 · 2026-08-22
- 'Harness Engineering' Rises: Custom Scaffolds Become the Foundation of AI-Native Companies — omarsar0 · 2026-08-22
- Bootstrapped to $1M+ in 18 Months: A Look at 40 AI Agents Running the Business — aryanXmahajan · 2026-08-22
- Codex Builds Working Circuits Inside the Game 'Turing Complete' — Full CPU Next — Angaisb_ · 2026-08-22
- 2000 multimodal patent project rebuilt in a few Grok prompts 26 years later — Daniel_Farinax · 2026-08-22
- OpenAI lets MCP plugins ship bundled "skills" baked into ChatGPT and Codex — dfinke · 2026-08-22