TraceLab provides real-world coding agent traces collected from Claude Code and Codex. The dataset captures multi-step agent sessions, LLM requests, tool calls, cache behavior, latency distributions, and workload characteristics for studying agentic AI systems.
TraceLab is integrated into NVIDIA AIPerf, which can replay TraceLab sessions to benchmark serving systems under real agentic coding workloads.
Explore the dataset on the project website, or read the paper and blog post.
Kan Zhu 朱侃
PhD Student (Computer Science)
I am Kan Zhu, a fourth year PhD student at University of Washington’s Paul G. Allen School of Computer Science and Engineering.