Coding agents burn through LLM infrastructure differently than anyone had measured before.
Researchers released TraceLab, a dataset of roughly 4,300 real coding-agent sessions drawn from day-to-day use of Claude Code and Codex. The trace logs about 350,000 LLM steps and 430,000 tool calls, making it the first public record of coding-agent traffic at this scale across multiple agents and model families. Prior benchmarks existed, but none captured the messy, continuous patterns of actual use — the kind of data you need to tune a serving system.
The findings complicate the assumption that coding agents are just "chatbots that also run code." The workloads feature long autonomous loops, contexts that grow large while outputs stay short, and tool-call distributions with heavy tails — meaning a small number of tool types dominate, but outliers are common enough to matter. Prefix cache hit rates are high but not high enough to ignore the gaps, particularly around human-paced pauses between sessions.
That last detail is the sharpest operational insight: when a developer steps away mid-task, cached key-value state decays in ways current systems do not account for well. The researchers point to four concrete optimizations — lower-overhead tool calling, append-length-aware prefill scheduling, semantic-aware tool-latency prediction, and smarter KV-cache management around those human gaps. None of these are exotic; they are engineering work that now has empirical justification rather than speculation.
As inference costs remain one of the loudest complaints from teams running agents at scale, a public workload trace is more useful than another benchmark promising to rank models on toy tasks. The dataset, pipeline, and analysis code are all open-source.