We are witnessing the dawn of AI automation at scale, where agent collaboration is redefining what’s possible in the workplace. And they are multiplying fast – bringing the promise of more automation, but also new risks: an expanded attack surface demanding robust agent identity verification, fragmented data requiring standardization to fuel agent intelligence, and the inherent complex interoperability of managing multi-agent systems where agents and specialized sub-agents must operate safely and effectively without losing a split second of performance.

While open protocols like model context protocol (MCP) and other open-source initiatives play a big role in bringing together industry stakeholders – much like how HTTP standardized web communication – enterprises, researchers, and open-source communities are increasingly rallying around standardization efforts that will accelerate the era of agentic AI.

By aligning on standardizations, we can expedite new levels of adoption making it easier and safer for networks to adapt and scale to a businesses’ needs. But rapid deployment of agentic AI and multi-agent systems has profound implications for enterprise network infrastructure and the architectural decisions required to sustain a reliable network.

The promise of multi-agent systems

Standardized agent-to-agent protocols are ushering in multi-agent systems where interconnected AI agents can collaborate to execute long, autonomous workflows that span multiple environments. These workflows are far more complex than traditional AI processes and involve continuous interaction between agents with different capabilities, roles, and locations.

Consider a health care scenario: a diagnostic agent on a hospital network could analyze patient scans using a local small language model (SLM), flagging anomalies. Simultaneously, it could collaborate with a cloud-based agent to compare results with a global dataset and generate treatment recommendations using a large language model (LLM). A logistics agent might then autonomously coordinate with pharmacy and medical equipment APIs to ensure timely delivery. This workflow spans local, cloud, and API systems, operating autonomously to reduce delays and improve patient outcomes.

Or a manufacturing scenario: a quality control agent at a factory floor could inspect product quality in real time using machine vision models that run locally. If defects are detected, it could engage a predictive maintenance agent to determine potential servicing, or a production optimization agent to adjust machine settings and communicate with other factories in the network to ensure consistency. Such a workflow involves tight coordination between local systems, cloud-based analytics, and inter-factory communication – all operating autonomously to minimize downtime and improve efficiency.

Similar transformations are emerging across retail and finance too, where agents collaborate across edge devices, on-premises infrastructure, and cloud services to optimize operations, enhance customer experiences, and improve efficiency.

Implications for network architecture

In multi-agent ecosystems, AI agents don’t just execute isolated tasks, they collaborate autonomously to achieve multi-step objectives such as:

• Task delegation where a general-purpose agent might break down a task into smaller subtasks, delegating them to specialized agents or tools
• Dynamic decision-making in which agents evaluate outcomes at each step and adjust the workflow dynamically based on real-time data or changing requirements
• Cross-environment communication where agents interact across environments to seamlessly blend on-premises and cloud intelligence, like SLMs in an office, cloud-based LLMs, and APIs.

These workflows can be persistent and autonomous, running for hours, days, or even longer without human intervention. For instance, an agent might monitor a manufacturing process continuously, adjusting operations in real time based on IoT sensor data, while collaborating with cloud-based analytics models to predict maintenance needs.

It's no surprise, then, that the autonomous and persistent communication between AI agents across local, cloud, and hybrid environments introduces new demands that challenge traditional architectures.

New infrastructure priorities for new demands

For IT leaders supporting the infrastructure for multi-agent systems, the challenge lies in building networks that can handle the scale, complexity, and security requirements of these next-generation AI workflows. New considerations include:

• Connectivity infrastructure to support high-bandwidth, low-latency communications. Many AI workflows, particularly in real-time applications like manufacturing or health care, require ultra-low-latency inter-agent communication with near-zero risk of downtime. This communication is often bidirectional, meaning networks previously optimized for download-heavy traffic now have to support symmetrical bandwidth.
• Edge computing capabilities to handle local processing needs, reduce WAN traffic, and improve response times. Local SLMs and edge devices require robust on-premises compute and networking resources, while cloud-based agents and APIs require high-capacity WANs. As such, network architectures now must be able to blend local and cloud-based intelligence to balance performance, cost, and privacy.
• End-to-end encryption, robust authentication, and AI governance frameworks to secure inter-agent communications. Autonomous inter-agent communication introduces new risks, such as unauthorized data sharing or malicious actors impersonating agents, putting new demands on zero-trust architectures, and governance policies.

Without a resilient, secure, and high-performance network, the promise of agentic AI will stay out of reach. Networks are the backbone that turn AI potential into reality.

Building the network infrastructure for the future, again

The network is the nervous system of the AI-driven workplace: it delivers every workload, every application, every connection, and every interaction in the digital world. IT has adapted to technology disruptions before. But now with AI agents at the core, collaborating nonstop, delegating tasks, making decisions, and crossing cloud and edge boundaries, these workflows can run for days and demand that networks never drop the ball. Time for networks to evolve again.