TraceLab is a trace of real coding-agent usage collected and released by the authors of TraceLab: Characterizing Coding Agent Workloads for LLM Serving, a University of Washington-led study, to support research on serving coding agents efficiently. It contains about 4,300 sessions comprising roughly 350,000 LLM steps and 430,000 tool calls, gathered from 43 developers over about eight months of their day-to-day use of Claude Code and OpenAI Codex, across more than 20 model versions. The authors describe it as, to their knowledge, the first large-scale cross-provider trace of real coding-agent use.
Contents
The trace is normalised into a step-level schema in which each row is one step: a single LLM invocation together with the tool calls it produces. Each step’s input is decomposed into prefix tokens (history read from the prefix cache) and append tokens (newly added input, including user messages, tool results and tokens introduced by cache misses), alongside output tokens, and each step keeps an ordered list of timestamped events — user messages, tool results, reasoning and output text, and tool calls — from which its timeline can be reconstructed.
By provider, the trace holds 2,676 Claude sessions from 37 users (October 2025 to June 2026) and 1,589 Codex sessions from 22 users (September 2025 to June 2026), for 357,161 LLM steps and 432,510 tool calls in total. It covers 23 models: within Claude, mostly opus-4-7 (63.1%), and within Codex, mostly gpt-5.5 (47.5%) and gpt-5.4 (26.0%). Total input amounts to 54.90 billion tokens, of which 52.56 billion are prefix tokens and 2.34 billion append tokens, with 186.9 million output tokens; the authors estimate its API list-price equivalent at about $40.4K.
Provenance
The data comes from the session logs that Claude Code and Codex persist by default, which record metadata, user messages, encrypted reasoning, intermediate output, tool calls and tool results. The authors built a pipeline that extracts these logs, reconciles the two agents’ differing event structures, token accounting and tool-timing metadata into the unified schema, and anonymises the result. To protect user privacy only a sanitised trace is released: session, tool-call, project and user identifiers are replaced with stable pseudonyms, and raw user messages and tool input/output text are dropped, keeping only their character counts along with token usage and timestamps. The dataset, the collection pipeline and the analysis code are published in the project’s GitHub repository, with a project website at https://tracelab.cs.washington.edu.
Because the sessions come from the authors’ own research workflows — system building, evaluation and analysis — the authors caution that results drawn from it may not fully generalise to other organisations, development workflows or agent domains.
Use
TraceLab: Characterizing Coding Agent Workloads for LLM Serving uses the trace to characterise coding-agent workloads for LLM serving, reporting among other findings a 95.7% global prefix-cache hit rate with misses concentrated after human-paced gaps, and prefix-token re-reads accounting for 59.5% of priced cost.