Alterion: governing agents across the network

Sanford Redlich is Head of Engineering at Alterion, a company that governs AI agents on large corporate systems. Luca Berton talked to him at AGNTCon + MCPCon North America 2026 in San Jose on 23 October 2026 and asked what Alterion does, and what sets it apart.

Three differentiators

The first is completeness across the network. “We cover everything,” he says, and the platform classifies many kinds of agentic behaviour that matter to a company: compliance, legal, return on investment. Luca asked whether that makes it an incident response tool. The system is good at incident response, he says, but its primary job is to observe agentic behaviour across the whole network, from the endpoints to the services outside, identify what it is and, if necessary, control it, warn about it or take some other action.

The second is that it runs on premises and can be air-gapped, so the data never has to leave the customer’s site. The third is a proprietary AI system for the classification. That matters because the agents being governed are non-deterministic, and a deterministic system will not keep up with them: “you have to have systems that can adapt as fast as the entities that you’re trying to control.”

How hard can an endpoint be

Take a laptop. Some agents there, the Claude app, Copilot, VS Code, have API control planes you can log into to see what they are doing. Many others offer none, so Alterion has to intercept their network traffic and decode how they communicate, well enough to recognise the behaviour. Then the agents call MCP servers on the network, which Alterion catalogues, and MCP servers on the same machine over standard IO, which have to be intercepted as well.

That is why the work needs engineers who understand systems design and system dynamics and are comfortable with the pace of agentic development: “we have to develop our systems faster than the systems that we’re trying to observe.”

Making it readable

The interface is very visual, partly for the volume. One company Alterion talks to has 10,000 agents of a single kind; there could easily be hundreds of thousands, each communicating its own way. The task is to classify all of it without too many mistakes, then show it in a form that makes sense to a CISO, a department head or whoever is responsible for compliance. “It’s an extreme challenge.”

Luca asked about Europe’s regulations. Sanford knows GDPR’s Article 9 and Article 22, and wants the classification accurate enough to identify violations of them. In practice, a leader sees an overview of the company or a team and the agents on its network. If the finance department seems to be doing something odd, an investigate button shows the traffic, who is doing what and which agentic workloads are running. Then a control can be drawn around a boundary, the finance department for instance, so that agents of a given type cannot send certain information. Social security numbers are the trivial example; the controls can get much more fine-grained.

A third kind of entity

Alterion works with healthcare, financial and big tech companies in the United States, and is beginning to work internationally. Regulation, he says, scares a lot of companies: “Nobody wants to be the CEO going in front of the board and explaining why regulations were not complied with by agents on your network.” The agents will be there regardless. People were unpredictable, but they went through HR; software was deterministic and did what it was asked. Agents are a third kind of entity, unpredictable and very fast.

How to reach Alterion

The way in is the website, alterion.ai, and a sales conversation. There is no download or installation for now, because Alterion is optimised for Fortune 100 and Fortune 500 customers, a set it plans to widen over time.

More

The episode with Sanford Redlich is being produced and goes out on YouTube first. The day it was recorded: AGNTCon + MCPCon North America 2026: day two.