ClawGym II：黑盒强化学习训练 Agent 框架与实验结果
You can now RL-train an agent through the same complex harness it will actually…
Agent 训练的新范式，黑盒 RL 方案极具参考价值，建议关注其技术细节。
You can now RL-train an agent through the same complex harness it will actually run in, without needing access to the harness internals.
你现在可以通过与实际运行环境相同的复杂框架来对智能体进行强化学习（RL）训练，而无需访问该框架的内部实现。
ClawGym II shows that Claude Code or OpenClaw can be treated as a black box and still become part of the RL training loop.
ClawGym II 表明，Claude Code 或 OpenClaw 可以被视为黑盒，并仍能成为强化学习训练循环的一部分。
The framework runs OpenClaw or Claude Code unchanged inside sandboxes, intercepts model calls at the serving boundary, and rebuilds fragmented calls into prefix-tree trajectories that PPO or GRPO can optimize.
该框架在沙箱内无修改地运行 OpenClaw 或 Claude Code，在服务边界拦截模型调用，并将碎片化的调用重建为 PPO 或 GRPO 可优化的前缀树轨迹。
That lets the model learn through the harness without the training stack reproducing its tool routing, retries, context management, or subagents.
这使得模型能够通过框架进行学习，而训练栈无需复现其工具路由、重试、上下文管理或子智能体机制。
With Qwen3-30A3B, this raised ClawGym-Bench Pass@1 by 9.98 points through OpenClaw and 14.81 points through Claude Code.
使用 Qwen3-30A3B，通过 OpenClaw 使 ClawGym-Bench Pass@1 提升了 9.98 分，通过 Claude Code 则提升了 14.81 分。
Mix-harness training also worked: a policy trained from OpenClaw and Claude Code matched or slightly beat the corresponding single-harness models under both execution systems.
混合框架训练同样有效：基于 OpenClaw 和 Claude Code 训练的策略，在两种执行系统下均达到或略优于对应的单一框架模型。
The paper also reports gains on JobBench and OfficeQA, so the setup extends beyond ClawGym-style tasks.
论文还报告了在 JobBench 和 OfficeQA 上的性能提升，说明该设置不仅限于 ClawGym 风格的任务。
– arxiv. org/abs/2608.16798
– arxiv.org/abs/2608.16798
Title: "ClawGym II: Exploring Black-Box RL on Agent Harness"
标题：《ClawGym II：探索智能体框架上的黑盒强化学习》
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力