Amazon Researchers Detail Eluna for Graph-Guided Warehouse Agents

Amazon researchers introduced Eluna, an agent system for executing warehouse standard operating procedures. The framework encodes each procedure as a directed acyclic graph, exposes only reachable branches and current node details, and assigns independent node evaluations to parallel sub-agents. The main agent retains graph-level state and action sequencing.

The paper also describes asymmetric episodic distillation. A strong teacher uses memories of prior errors to improve training trajectories; a smaller student is fine-tuned on corrected trajectories after the memory is removed. The intended benefit is to internalize corrections without paying the teacher's inference-time latency and memory overhead.

On a 13-task benchmark and two production applications, the authors report that fine-tuned models match or exceed larger baselines and achieve 94% expert agreement on ticket processing. The manuscript is Amazon-authored and has not been independently reproduced. Expert agreement in one application is not universal warehouse accuracy or proof that every operational action is safe.

Eluna addresses two common agent failures: context overload and procedural drift. A full procedure can overwhelm a model with irrelevant branches, while a free-form agent may skip mandatory checks. Progressive disclosure narrows context; a graph constrains reachable actions; isolated sub-agents prevent tool output from flooding the coordinator.

A production evaluation should replay known incidents and inject ambiguous states, missing telemetry, tool timeouts, conflicting measurements, and mid-run procedure changes. Reviewers should measure node-level accuracy, exact path compliance, action correctness, escalation rate, recovery behavior, latency, and cost.

The graph itself becomes production code. It needs version control, approval, test fixtures, rollout gates, and a mapping from every action to authorization and rollback. Model evaluation cannot compensate for a procedure that is incomplete or outdated.

LDS sees Eluna's strongest idea as separating orchestration structure from model capability. The system makes procedure state inspectable and limits context by construction. Its main evidence gap is external validity: the public paper does not allow independent teams to confirm performance across other warehouses, tools, procedures, and failure modes.

Watch for code or benchmark release, independent reproduction, failure-case reporting, procedure-update tests, authorization controls, cost comparisons, and evidence that smaller students retain rare but critical branches.