TraceLab: 4,300 real coding-agent sessions reveal high cache hit rates but quadratic cost growth

CShorten30 · x · 2026-08-19

A new paper TraceLab offers a large-scale characterization of real coding-agent workloads.

The authors collected and released traces from their own day-to-day use of Claude Code and Codex: 4,300 coding-agent sessions, 350K LLM steps and 430K tool calls — filling a gap left by public traces that fail to capture real multi-agent, multi-model serving patterns.
Key findings: coding-agent workloads feature long autonomous loops, long contexts with short outputs, heavily-tailed tool calls, and high but imperfect prefix cache hit rates.
Concrete serving opportunities: lower-overhead tool calling, append-length-aware prefill, semantic-aware tool-latency prediction, and better KV-cache management around human-paced gaps.

The quoted caching math adds: after a coffee break, a simple "hi" can cost a full dollar — your cache went cold, so you repay for the million tokens of context; even with caching, cost grows quadratically with conversation length, which is why Claude Code/Codex compact context around 200-300K tokens.

Original post →

TraceLab: 4,300 real coding-agent sessions reveal high cache hit rates but quadratic cost growth

More from coding & agent