Autonomous Event-Driven Multi-Agent Orchestration for Enterprise AI at Scale

This research addresses the challenge of scaling multi-agent AI systems for enterprise environments. While many current systems rely on simple request-response workflows, true enterprise AI requires continuous monitoring, detection, and action across large groups of specialist agents. The authors investigate how existing orchestration architectures perform as the number of agents increases and introduce a new "Task Manager" to handle the complexities of large-scale, event-driven operations.

Evaluating Orchestration at Scale

The researchers tested two popular multi-agent architectures—DAG Plan and Execute and ReAct—across 208 production-derived scenarios. These scenarios ranged from small "Persona" groups (fewer than 10 agents) to "Department" levels (20–80 agents) and full "Enterprise" scales (200 agents). The study found that the primary factor limiting performance is the total scale of the system rather than the complexity of the individual tasks. As the number of agents grows, "agent discovery noise" becomes a significant bottleneck, causing performance to degrade. Interestingly, simple tasks were found to suffer more from this degradation than complex ones.

Comparing Architectures

The study highlights distinct trade-offs between the two tested architectures:

• DAG Plan and Execute: This approach excels at smaller scales by offering higher precision and structured parallelization. However, its operational overhead becomes a liability as the system scales up to enterprise levels.

• ReAct: This architecture proves to be more robust at larger scales because it handles failures incrementally, making it more resilient than the structured DAG approach when managing many agents.

The Role of the Task Manager

To address the limitations of existing systems, the authors introduced a Task Manager designed for continuous operation. This component manages the flow of work through three primary mechanisms: priority inference, related-event merging, and preemption. By implementing this manager, the researchers observed significant improvements at the enterprise scale, including a 14–75% reduction in high-priority queue latency and a 20 percentage point increase in the correctness of related-event handling.

Key Takeaways

The research demonstrates that moving toward continuous, event-driven enterprise AI requires moving beyond simple request-response models. While existing architectures like DAG Plan and Execute and ReAct are effective for smaller teams, they require additional infrastructure—such as a dedicated Task Manager—to remain functional and efficient when deployed at the scale of an entire enterprise.