Eluna: An Agentic LLM System for Automating Warehouse Operations

Jul 16, 2026

Upload by Rowan

Warehouse operations are governed by Standard Operating Procedures (SOPs) that specify complex, multi-system decision logic and must be executed reliably under tight time constraints. LLM agents struggle with enforcing procedural compliance and suffer from context overload when full SOPs are provided. Eluna is a production-deployed agentic framework that encodes SOPs as directed acyclic graphs with progressive disclosure and parallel sub-agent task execution using persistent code execution and live data access.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

Introduction

SOP-guided warehouse workflows and challenges

Eluna: graph-guided agent framework

Trajectory-centric training with asymmetric episodic distillation

Contributions and real-world results

Eluna: An Agentic LLM System for Automating Warehouse Operations with Reasoning and Task Execution Ning Liu* Kalle Kujanpää* Zhaoxuan Zhu* P Aditya Sreekar* Kaiwen Liu Chuanneng Sun Jorge Marchena Menendez Matthew Bales Tianyu Yang Shahnawaz Alam Rose Yu Baoyuan Liu Kristina Klinkner Shervin Malmasi [Amazon.com](Amazon.com), Inc. Fulfillment Technologies and Robotics {ningliun,[malmasi}@amazon.com](malmasi}@amazon.com) arXiv :2607 .08960v 1 [ cs .LG] 9 Jul 2026 Abstract Warehouse operations are governed by Standard Operating Procedures (SOPs) that encode complex, multi-system decision logic, which must be executed reliably under strict time constraints, yet LLM agents lack mechanisms to enforce procedural compliance and degrade under the context overload full SOP specifications introduce. We present Eluna, a production-deployed agentic system for reliable SOP execution. Eluna is a graph-guided, multi-agent framework that encodes SOPs as directed acyclic graphs with progressive disclosure and delegates independent tasks to parallel sub-agents, each with persistent code execution and live data access. To meet production latency and accuracy needs, we use asymmetric episodic distillation where a strong teacher is improved through episodic error memories, then a smaller student is fine-tuned on the corrected trajectories with memory stripped, internalizing corrections without inference-time overhead. On a 13-task benchmark and two production applications, our fine-tuned models match or exceed their teacher, beat all larger off-the-shelf baselines, and reach 94% expert agreement on the ticket processing application. 1 Introduction Modern warehouse operations require continuous monitoring, multi-step diagnosis, and timely intervention across dozens of systems, ranging from robotic conveyance operators tracing threshold breaches to root causes within minutes, to inventory ticket processing that queries state, validates constraints, and submits physical item picks across many systems per ticket. Robotic conveyance diagnosis and ticket processing are two among many such workflows, each governed by Standard Operating Procedures (SOPs): directed decision pathways with prescribed steps, dependencies, quantitative thresholds, and constraints that must be followed faithfully. Today much of their execution is manual, * Equal contribution. slow, and error-prone, and delays in diagnosis directly degrade throughput, making automated SOP execution a pressing business need. Large language models reason well (Wei et al., 2022 ; Kojima et al., 2022), but agentic frameworks such as ReAct (Yao et al., 2023) and Reflexion (Shinn et al., 2024) rely on loosely structured prompts with no mechanism to strongly guide SOP compliance, so a capable model can still deviate from prescribed decision paths. Recent benchmarks confirm this gap on workflow-guided tasks (Xiao et al., 2024 ; Wang et al., 2025), industrial SOP-following (Nandi et al., 2025), and operational diagnostics where even SOP-enhanced multi-agent systems plateau in performance (Pei et al., 2025) . A core bottleneck is context overload: with the full SOP in view, performance degrades as complex workflows overwhelm the model’s ability to select relevant actions (Xiao et al., 2024 ; Pei et al., 2025) . This plateau holds regardless of model scale, and prompting alone is insufficient without domain-specific training (Nandi et al., 2025) . We present a graph-guided agent framework that addresses these challenges through a unified execution model. SOPs are encoded as directed acyclic graphs, and progressive disclosure surfaces only the reachable subgraph and node-level specifications on demand. A main agent orchestrates traversal and delegates independent node evaluations to parallel sub-agents in isolated contexts, each with a persistent code interpreter (the CodeAct paradigm of Wang et al. (2024)) and access to real-time data via the Model Context Protocol (Anthropic, 2024

FAQs

What problem does Eluna address in warehouse operations?

Eluna targets reliable execution of SOP-driven, multi-system workflows under strict time constraints, where current LLM agents cannot consistently enforce procedural compliance.

How does Eluna enforce SOP compliance during execution?

Eluna encodes SOPs as directed acyclic graphs and uses progressive disclosure so agents only see reachable portions of the workflow while delegating independent evaluations to parallel sub-agents.

What training approach does Eluna use to improve performance without inference-time episodic memory?

Eluna applies asymmetric episodic distillation: a strong teacher improves via episodic error memories, while a smaller student is fine-tuned on corrected trajectories with episodic memory stripped, internalizing corrections in weights.

Jul 16, 2026

Upload by Rowan