SemiAnalysis publishes AgentX from 393 Claude Code sessions

SemiAnalysis has published AgentX v1.0, a replay benchmark that turns 393 opt-in Claude Code sessions into a reproducible workload for stress-testing inference engines. The median session sends 142k input tokens and 444 output tokens per request, and 44% use subagents.

To make the traces shareable without leaking user prompts or code, the methodology page reduces inputs to what it calls "session-scoped chained hashes in 64-token blocks, preserving matching prefixes without revealing content." Sessions had to log at least 20 requests, run on Claude Code 2.1.139 or newer, and keep concurrent subagents under 10. Replay reconstructs the traffic into directed acyclic graphs, warms each lane with 10 requests, and profiles for one hour with the conversation held between 25 and 75 percent full.

The page frames the point of the exercise in blunt terms: "The most useful thing AgentX produced in its first months was not the open-source datasets. It was 50+ upstream pull requests from AgentX partners, optimizing real-world agentic workloads." Named partners include vLLM, SGLang, TensorRT-LLM, NVIDIA Dynamo, AMD ATOM, ROCm AITER, LMCache and Mooncake; SemiAnalysis lists OpenAI, Meta, Oracle and Microsoft among trusted users of InferenceX.