OPEN LINK ↗
// 2h agoRESEARCH PAPER
ClawGym II Makes Black-Box Agent RL Practical
ClawGym II introduces a framework for training agents through opaque harnesses such as OpenClaw and Claude Code, capturing model calls through a serving proxy and optimizing reconstructed trajectories. On Qwen3-30A3B, it improves ClawGym-Bench Pass@1 by up to 14.81 points.
// ANALYSIS
The paper’s strongest idea is treating production-grade agent harnesses as trainable infrastructure rather than rebuilding simplified agent loops for every experiment.
- –Prefix-tree trajectory reconstruction makes multi-turn, branching harness behavior compatible with PPO and GRPO
- –Sandbox isolation enables concurrent rollouts while preserving realistic workspace and tool interactions
- –Mix-harness training suggests one policy can learn across substantially different agent runtimes
- –Gains transfer beyond ClawGym-Bench to PinchBench, JobBench, and OfficeQA
- –The black-box framing improves practical relevance, but captured-call observability and verifier quality remain important constraints
// TAGS
clawgym-iiagenttrainingreinforcement-learningtool-useevaluationtraining-infraopen-source
DISCOVERED
2h ago
2026-08-18
PUBLISHED
2h ago
2026-08-18
RELEVANCE
9/ 10
AUTHOR
omarsar0