TraceLab: Characterizing Coding Agent Workloads for LLM Serving

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Kan Zhu, Mathew Jacob, Chenxi Ma, Yi Pan, Stephanie Wang, Arvind Krishnamurthy, Baris Kasikci
• Venue: arXiv: cs.LG
• Published: 2026-06-29
• Collected: 2026-06-30
• URL: http://arxiv.org/abs/2606.30560v1
• PDF: https://arxiv.org/pdf/2606.30560v1
• DOI: UNKNOWN

Plain-Language Brief

Researchers have collected and released a dataset of 4300 coding agent sessions, containing around 350,000 LLM steps and 430,000 tool calls from their own use of Claude Code and Codex. The analysis shows that coding-agent workloads have long autonomous loops, long contexts with short outputs, diverse and heavily-tailed tool calls, and high but imperfect prefix cache hit rates.

What the Document Says

• The paper discusses the challenges in serving coding agents efficiently using LLMs. It mentions that there is a lack of data for analysis of real workload patterns in this area. The authors have collected a trace of approximately 4300 coding-agent sessions from their own use of Claude Code and Codex, containing about 350,000 LLM steps and 430,000 tool calls. The analysis reveals that these workloads consist of long autonomous loops, long contexts with short outputs, diverse and heavily-tailed tool calls, and high but imperfect prefix cache hit rates.

Key Concepts Mentioned

• coding agents
• LLMs
• serving efficiency
• workload patterns
• data analysis
• Claude Code
• Codex
• long autonomous loops
• long contexts with short outputs
• diverse tool calls
• prefix cache hit rates

Claims Made by the Document

• There is a lack of data for understanding real workload patterns in coding agent usage.
• The authors have collected a dataset of approximately 4300 coding-agent sessions.

Evidence Used

• The text mentions that existing public traces and benchmarks do not capture real, day-to-day coding-agent usage across multiple agents and model families for serving-system analysis.
• The text states that the authors have collected a trace of roughly 4300 coding-agent sessions.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Coding agents are rapidly becoming a major application of agentic LLMs, but serving them efficiently remains challenging. Progress on this challenge requires understanding real workload patterns, yet the data needed for such analysis is largely absent. Existing public traces and benchmarks do not capture real, day-to-day coding-agent usage across multiple agents and model families for serving-system analysis. To help fill this gap, we collect and release a trace of roughly 4,300 coding-agent sessions, containing about 350,000 LLM steps and 430,000 tool calls from our own day-to-day use of Claude Code and Codex. Our analysis shows that coding-agent workloads feature long autonomous loops, long contexts with short outputs, diverse and heavily-tailed tool calls, and high but imperfect prefix cache hit rates. These findings point to concrete opportunities for optimizing serving, including lower-overhead tool calling, append-length-aware prefill, semantic-aware tool-latency prediction, and improved KV-cache management around human-paced gaps. We release the dataset, trace collection pipeline, and analysis code at https://github.com/uw-syfi/TraceLab.git; the project website is https://tracelab.cs.washington.edu.