Multi-Agent Orchestration for Enterprise AI: arXiv Paper

Harsh Rao Dhanyamraju, Leonidas Raghav and Aaron Lee submitted an arXiv paper on 18 Jun 2026 titled "Autonomous Event-Driven Multi-Agent Orchestration for Enterprise AI at Scale." The paper evaluates DAG Plan and Execute and ReAct across 208 production-derived enterprise scenarios spanning Persona (<10 agents), Department (20-80), and Enterprise (200) scales, and introduces a Task Manager for continuous operation via priority inference, related-event merging, and preemption.

What did the authors test and how?

The paper tested two orchestration architectures, DAG Plan and Execute and ReAct, across 208 production-derived scenarios at three scale tiers: Persona (<10 agents), Department (20-80), and Enterprise (200). The evaluation framework treats enterprise AI as continuous event monitoring, detection, and action across specialist agents rather than discrete request-response workflows, and it measures how each architecture behaves as the number of agents and discovery noise grow.

The authors also introduce a Task Manager designed for continuous operation, which performs priority inference, related-event merging, and preemption to manage queues and related events during runtime.

How did DAG Plan and Execute compare with ReAct?

DAG Plan and Execute delivered higher precision and more structured parallelization at smaller scales, but its higher coordination overhead worsened performance at enterprise scale; ReAct proved more robust by handling failures incrementally. Both architectures performed well at small scale, the paper finds, but degraded at enterprise scale where agent discovery noise becomes the primary bottleneck.

The study highlights an unexpected pattern: simple tasks degraded more sharply than complex ones as scale increased, indicating that scale, not task complexity, dominates orchestration performance in these production-derived scenarios.

What concrete effects did the Task Manager produce?

At enterprise scale the Task Manager reduced high-priority queue latency by between 14% and 75% and improved related-event correctness by over 20 percentage points. The Task Manager’s combination of priority inference, related-event merging, and preemption enabled more continuous operation under large-scale noise and discovery churn than either orchestration architecture alone.

Those quantitative results are the primary measured improvements reported for enterprise-scale scenarios in the paper.

Why it matters

Enterprise deployments commonly multiply the number of specialist agents and the volume of events. The paper shows that adding agents changes the dominant failure mode: agent discovery noise, not task complexity, becomes the limiting factor. That shifts engineering priorities toward discovery robustness, queue management and preemption logic. The Task Manager’s measured reductions in latency and gains in related-event correctness point to practical mitigations operators can deploy without replacing their orchestration approach.

What to watch

Watch for follow-up evaluations that replicate these enterprise-scale conditions beyond the 208 production-derived scenarios and for open-source or vendor implementations of the Task Manager’s priority inference and preemption features. Also track whether subsequent work measures the trade-off DAG-style coordination imposes at very large agent counts versus ReAct’s incremental failure handling.

Paper and metadata: arXiv:2606.20058, submitted 18 Jun 2026; authors Harsh Rao Dhanyamraju, Leonidas Raghav, Aaron Lee.