Which harnesses were used for the benchmarks? | 这些基准测试使用了哪些 Agent Harness？
Hi StepFun team, thank you so much for sharing such impressive models and research!
A quick question about reproducing the Step 5 Preview benchmark results: which agentic harnesses (eg, Step Code, Claude Code, Codex, or custom setups) were used for ProgramBench, Agents' Last Exam (ALE-CLI), FrontierFinance, Terminal-Bench 4.0, and DRACO? Was Step Code used for any of them?
I couldn't find a benchmark-to-harness mapping or detailed evaluation methodology. Any details you could share would be greatly appreciated by the research community! 🙏
StepFun 团队你们好！非常感谢你们分享如此出色的模型和研究成果！
想请教一个关于复现 Step 5 Preview 基准测试结果的问题：在 ProgramBench、Agents' Last Exam (ALE-CLI)、FrontierFinance、Terminal-Bench 4.0 和 DRACO 的评测中，你们使用了哪些 Agent Harness（例如 Step Code、Claude Code、Codex 或自定义框架）？其中是否有使用 Step Code？
我没有找到各基准测试对应的 Harness 配置或详细评测方法。如果能分享一些相关的复现细节，相信会对研究社区非常有帮助！非常感谢！🙏