Agentic AI Workload Characteristics

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Yuan, Yichao, Nayak, Ankita, Kundu, Souvik, Talati, Nishil
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866913162439163904
author Yuan, Yichao
Nayak, Ankita
Kundu, Souvik
Talati, Nishil
author_facet Yuan, Yichao
Nayak, Ankita
Kundu, Souvik
Talati, Nishil
contents Agentic AI shifts LLM serving from isolated prompt-generation requests to stateful, multi-turn executions that repeatedly invoke the model, call tools, and grow context over time. This paper characterizes ReAct-style agents from both the LLM-serving and tool-execution perspectives using an end-to-end tracing infrastructure across reasoning and non-reasoning Gemma and Qwen configurations on five agentic benchmarks. Our study shows that agentic workloads are not simply long-prompt workloads: with effective context caching, most input tokens are reused across turns, making execution decode-dominated while increasing dependence on long-lived KV-cache state. We also find that tool use has a clear temporal structure, with agents shifting from read/explore behavior early in execution to execute/write behavior later. These results show that efficient agentic serving must jointly manage repeated model re-entry, persistent context state, and workload-dependent tool behavior.
format Preprint
id arxiv_https___arxiv_org_abs_2605_26297
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Agentic AI Workload Characteristics
Yuan, Yichao
Nayak, Ankita
Kundu, Souvik
Talati, Nishil
Distributed, Parallel, and Cluster Computing
Agentic AI shifts LLM serving from isolated prompt-generation requests to stateful, multi-turn executions that repeatedly invoke the model, call tools, and grow context over time. This paper characterizes ReAct-style agents from both the LLM-serving and tool-execution perspectives using an end-to-end tracing infrastructure across reasoning and non-reasoning Gemma and Qwen configurations on five agentic benchmarks. Our study shows that agentic workloads are not simply long-prompt workloads: with effective context caching, most input tokens are reused across turns, making execution decode-dominated while increasing dependence on long-lived KV-cache state. We also find that tool use has a clear temporal structure, with agents shifting from read/explore behavior early in execution to execute/write behavior later. These results show that efficient agentic serving must jointly manage repeated model re-entry, persistent context state, and workload-dependent tool behavior.
title Agentic AI Workload Characteristics
topic Distributed, Parallel, and Cluster Computing
url https://arxiv.org/abs/2605.26297