Scaling Coding-Agent RL to 32x H100s. **Achieving 160% improvement** on Stanford's TerminalBench - View it on GitHub