Upper portion of the AgentStage accepted manuscript, showing its title, authors, and abstract.

AgentStage: Exploiting LLM Thinking for Data Staging in Scientific Agents

Izzet Yildirim1   Xian-He Sun1   Anthony Kougkas1

1Illinois Institute of Technology

IEEE eScience 2026 · Accepted

Paper coming soonCodeBibTeX

© 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.

Abstract

Scientific LLM agents spend much of a session inside tool calls reading data from storage. AgentStage uses the preceding thinking phase as a prefetch window: it reads streaming reasoning and filesystem probes to infer the next working set, stages files into local storage, and redirects tool reads to the hot copy when the tool fires.

Across curated scientific tasks and community benchmarks, AgentStage delivers an average 1.74× end-to-end speedup on curated tasks (up to 2.34× per run), with no per-model or per-task tuning.

Overview

The agent’s plan surfaces before its next tool call. AgentStage turns that idle storage interval into useful work.

LLM thinkingstream reasoning
Detect intentmatch files and probes
Stage datacopy cold to hot storage
Tool reads hottransparent path redirect

The capture proxy, tiered detector, and staging daemon run during the thinking phase. An LD_PRELOAD shim redirects reads only after the tool fires.

Results

AgentStage improves end-to-end sessions across different reasoning models and explains when staging will help before it runs.

End-to-end speedup

Mean curated-task speedup by reasoning model AgentStage speeds up curated tasks by 1.52 to 1.88 times on average across four reasoning models. 2.5× 2.0× 1.5× 1.0× MEAN SESSION SPEEDUP · 3 CURATED TASKS 1.73× Haiku 1.52× Sonnet 1.83× Flash 1.88× Qwen3 REASONING MODEL

Mean speedup over three curated scientific tasks. Values are computed from three replayed sessions per task and model.

Working-set detection

Tier-1 byte recall by reasoning model Across the curated evaluation sessions, AgentStage achieves 74 to 100 percent mean tier-1 byte recall depending on model. MEAN TIER-1 BYTE RECALL · CURATED SESSIONS 0%50%100% Haiku 95.2% Sonnet 99% Flash 73.8% Qwen3 100%

Mean tier-1 byte recall across curated sessions. The detector combines partial intent in the reasoning stream with filesystem-probe feedback.

Where the time goes

Session wall-time decomposition with and without AgentStage A representative curated session takes 92.43 seconds without AgentStage and 51.64 seconds with it. The 29.55 second thinking phase stays constant while tool time drops from 62.88 to 22.09 seconds. The control path is schematic: detection starts with thinking and dispatches staging before tool fire. REPRESENTATIVE CURATED SESSION · WALL TIME 0 s 50 s 92 s Naive AgentStage Thinking 29.55 s Tool 62.88 s Thinking 29.55 s Tool 22.09 s 92.43 s 51.64 s Detection Staging 40.79 s of tool-path time saved

The thinking interval stays fixed. Detection begins with the reasoning stream and dispatches staging to local hot storage before tool fire.

When does staging help?

The two conditions that predict when AgentStage helps A workload benefits most when cold I/O is a large share of its session and staging finishes before the tool fires. LOW I/O SHAREHIGH I/O SHARE COLD-READ SHARE OF SESSION WALL TIME PREFETCH FINISHES WITHIN THINKING WINDOW Compute-bound~1.03× ceiling Metadata-boundhigh I/O, misses deadline Bandwidth-bound1.74× averageprefetch completes yesno

Benefit depends on cold-read share and whether prefetch completes within the thinking window. These conditions can be estimated before staging.

Reproducibility

The public repository contains the AgentStage runtime, unit tests, claim-verification suite, and the figure-building pipeline. Raw run artifacts and datasets are not deployed with this site.

View code on GitHub

BibTeX

@inproceedings{yildirim2026agentstage,
  title = {AgentStage: Exploiting LLM Thinking for Data Staging in Scientific Agents},
  author = {Yildirim, Izzet and Sun, Xian-He and Kougkas, Anthony},
  booktitle = {IEEE International Conference on e-Science},
  year = {2026}
}