The AWS outage that took down a sizable portion of the web in October was caused by something almost mundane: automated systems that tried to update the same DNS record simultaneously, resulting in chaos that may have cost hundreds of millions of dollars.

And just a few weeks later, a similar such issue arose from Cloudflare, with the vendor’s machine learning-based Bot Management system falling foul of one of its own database systems – which resulted in Cloudflare’s servers “panicking” and eventually taking out yet more of the web.

For Jeff Gray, CEO of network automation vendor Gluware, these incidents illustrate a problem the industry isn't talking about enough. With enterprises racing to automate with agentic AI lathered across their infrastructure stack, who's making sure those agents don't break things at scale?

“If everyone's building an agent, then who's going to validate that all these different agents aren't just going to break things at scale?” Gray told SDxCentral. “If you look at the AWS outage in terms of DNS, some of these cascading outages were dramatic. There's got to be a layer of trust.”

That's the pitch behind Titan, Gluware’s new AI-powered network automation platform unveiled earlier this week.

Rather than positioning itself as yet another entrant in the already crowded AI-augmented network automation space, Gluware is betting that what enterprises actually need is a governed middleware layer that sits between the proliferating ecosystem of third-party agents and the multi-vendor infrastructure they're trying to manage.

Central to the new offering is Gluware's implementation of model context protocol (MCP), the open standard released by Anthropic that’s quickly become the default way for AI agents to connect to external systems.

But while the team at Gluware leans heavily on MCP, Gray argued that, alone, the protocol isn’t enough.

“You want some governance, you want some compliance, you want to make sure that the agents aren't just going to do their own thing and break the network at scale,” he said. “We took a little different approach, which is, let's create a framework that is trustworthy, where Gluware is in between the multi-vendor infrastructure and all the agents out of the box.”

Gray contrasted Titan with Cisco's AI Canvas, which he described as “very powerful for Cisco infrastructure only” while focused on surfacing problems rather than taking autonomous action.

“It's more about showing what's wrong, and then the humans need to go out and fix it,” he said. “Monitoring and observability platforms don't actually go off and take action and change things. What we've built with Titan is a framework where you can plug in third-party agents that can see things, and Gluware Titan is the brain that knows how to go and take action in a safe and predictable manner.”

Among those early integrations for Titan are NetBox and ServiceNow, enabling users to employ those systems' agents to coordinate with network automation workflows through Gluware's governed layer.

The CEO described it as an “Intelligent MCP Server” that effectively wraps the protocol with its existing Device Interaction Automation Layer (DIAL) technology to perform pre-checks and post-checks on any changes before they hit production infrastructure.

Gluware’s DIAL component is essentially a validation engine that helps perform those checks, which Gray told SDxCentral was built using over a million development hours, with almost 305 releases.

“DIAL’s device interaction automation layer is the commercialization of all that device interactivity, the two-way talk with safe and predictable, and pre-checks and post-checks,” Gray said. “If you can build automation 100 times faster, and we've already been getting 50-to-1 time and cost savings in our platform without AI, that’s an order of magnitude faster that we're going to see because it's already validated.”

While Gray spoke of some impressive time and cost savings in production deployments, he was careful not to overstate results from its current Early Field Trials, which included unnamed customers in sports and entertainment, healthcare, and engineering.

"I want to make sure that we have customers that stand up and say, 'I've been able to do this 500 times faster,'" Gray said. "But we're just not there yet."

General availability is planned for early 2026.