Skills
Black-box RL through an unmodified agent harness lifts Pass@1 by 9.98-14.81 points, with Claude Code as one of the harnesses
ClawGym II optimizes a policy through complex harnesses without instrumenting them: a serving proxy at the model boundary captures interactions, multi-turn trajectories are organized into a prefix tree, and PPO/GRPO are adapted for tree-structured optimization inside sandboxed execution. On Qwen3-30A3B it gained 9.98 points via OpenClaw and 14.81 via Claude Code on ClawGym-Bench, stable across 200-400 steps, with transfer to JobBench and OfficeQA. The structural idea worth stealing even without RL is treating the harness as a black box behind a proxy and reconstructing branching trajectories as a tree rather than a flat log.
↳ Follow the thread