Zayo Agentic Networking Opens Controlled MCP Actions

Zayo Agentic Networking adds a production networking-focused Model Context Protocol server to the DynamicLink platform, letting approved enterprise AI agents read network context and call authorised tools. Zayo made the feature available on September 8, one year after introducing DynamicLink as a software-controlled network service.

What Zayo Agentic Networking adds

DynamicLink already gave customers a software interface for network visibility and control. The new layer moves selected capabilities beyond the carrier portal, allowing an enterprise agent to gather network context or invoke a permitted action as one step inside a broader workflow.

Zayo says retrieval-augmented generation grounds interactions in DynamicLink knowledge, while its specialised networking skill and MCP server expose only the resources an organisation has authorised. Potential tasks include establishing cloud-to-cloud connectivity, investigating performance, coordinating multi-cloud changes and responding to unusual activity.

Zayo Agentic Networking turns network operations into controlled tools that an approved AI agent can call. Its value depends less on conversational convenience than on preserving the exact target, permission, execution result and rollback path across system boundaries.

Component	Role
DynamicLink	Network service, inventory and control plane
Networking skill	Domain instructions and workflow context
MCP server	Bounded interface to knowledge and authorised tools
Enterprise policy	Defines who or what can invoke each action
Human oversight	Approves sensitive changes and reviews outcomes

Independent reporting confirms the launch

BTW Media reported that the September 8 update opens DynamicLink information and authorised tools to agents outside Zayo’s portal, highlighting the need to preserve approval through execution. Converge Digest separately described the production networking MCP server and its governed access to context and actions.

Those reports confirm the mechanism and availability. Claims that the server is the industry’s first production MCP server built specifically for networking originate with Zayo and should be treated as a vendor characterisation rather than an independently measured market fact.

How enterprises should test it

Start with a read-only task: have an agent gather circuit context and propose a change without executing it. Compare the answer with the source inventory, check which fields were withheld and confirm that the agent cannot call an unapproved tool by changing the wording of its request.

Next, allow one reversible action in a sandbox or low-risk environment. The audit record should connect requester identity, policy decision, exact parameters, change window, result and rollback. This is the same evidence-first principle seen in how Google frames agentic AI threat response and in the human-review design discussed around the Observe.AI performance agents rollout.

FAQs

What is Zayo Agentic Networking?

It is a DynamicLink capability that gives approved enterprise AI agents governed access to network knowledge and authorised operational tools through MCP.

Can an agent make network changes?

Zayo says agents can invoke approved actions when an organisation exposes those tools. Customers define the available information, permissions, automation and human oversight.

Is the feature available now?

Zayo said Agentic Networking became immediately available as part of DynamicLink on September 8, 2026.

Designing a safe permission model

Permission should attach to a specific tool, network scope and change type rather than to a broad label such as trusted agent. A troubleshooting agent may need read access across several circuits but no ability to provision capacity. A provisioning agent may act only on pre-approved cloud connections during a defined change window.

Credentials also need short lifetimes and a clear service identity. If the same token can be reused outside the original workflow, the MCP boundary has not contained the risk. Logs should show the human sponsor, agent identity, tool version and policy rule that authorised every call.

Failure handling deserves its own exercise. Engineers should interrupt a workflow after the network action but before the agent reports success, then check whether DynamicLink and the enterprise system reconcile the actual state. They should also test duplicate requests, stale approvals and a rollback when only part of a multi-step change completes.

Finally, teams should measure operational improvement with modest metrics: time to assemble context, number of manual handoffs, failed changes and mean time to restore. The product will be valuable if it reduces coordination work while keeping authority legible, not simply because an AI agent can call a networking tool.