Runtime governance for AI agents

Why this conversation

Once agents act on behalf of people, someone will ask what they did and who let them. That question lands on the teams that run the platform, and in Europe it arrives alongside new regulation. At AGNTCon + MCPCon Europe in Amsterdam the host sat down with Sudeep Goswami, CEO of Traefik Labs, in the week Traefik launched an evidence layer for the decisions agents make.

The story of the week it was filmed is in AGNTCon + MCPCon Europe 2026: who we met.

Runtime governance and the evidence layer

Traefik describes itself as a cloud native and Kubernetes native platform. In an agentic workflow it adds the governance layer, and by governance Sudeep means runtime governance: the ability to enforce policies, whether a routing policy or a security policy, to observe everything through OpenTelemetry, and to record every decision the gateway made in what Traefik calls the evidence layer.

Auditing, he says, has been the missing layer in this workflow. The evidence layer is part of what Traefik calls the sovereign trust plane, launched earlier that week. It is designed to answer a few questions. What did the agent actually do? Who authorised the transaction, which may be an agent acting for another agent, or for a human, or any other combination? And can you prove what the agent did, both what it was allowed to do and what it was denied, in a verifiable way that flags any tampering? Those are the questions an audit will always ask.

Observability on its own would not be enough. It is sampled data and is meant for troubleshooting. The evidence layer records every transaction, both the allow decisions and the deny decisions, so that when an auditor asks for proof, it is there.

Sudeep names three groups who will consume that evidence. The security leadership, the CISO and the security team, want accountability and the ability to go back and verify. The governance, risk and compliance team wants to show that the controls actually work. And the product teams building the agents want to know what is working and what is not, so they can feed it into product development.

Open across the ecosystem

The platform is designed as a flexible enforcement point that stays open across several dimensions. It is agnostic to the agent and to the agent harness. It is agnostic to the identity platform, which matters because every agent now needs an identity: it works with OAuth vendors that support the standard protocols, among them OAuth 2.1, RFC 8693, ID-JAG and cross-app access, and emerging standards such as AuthZEN. At the observability layer it works with any vendor that supports OpenTelemetry.

On the access side, it can sit in front of any model exposed as an API endpoint, any MCP server that speaks the MCP protocol, and any API that follows the OpenAPI standards. Sudeep’s summary is a unified control plane through which an agent talks to models, to MCP servers through tool calls, and to APIs. Because the gateway is the entry point, it sees all of these conversations; once you can see them you can govern them, observability comes out of that, and the evidence layer records what happened.

One binary on Kubernetes

Everything comes pre-assembled in one binary. It is a paid offering, licensed with a token, and installs in minutes. It runs natively on Kubernetes, OpenShift included: as long as a distribution conforms to the upstream Kubernetes standards, it works out of the box, with no layer in between.

At the booth, a diagram showed a typical customer support workflow. Maya, a customer support representative, works with her agents on a ticket that asks for a refund to a customer. The diagram follows the authorisation of that refund from left to right, with Traefik Hub mediating, governing and observing the transaction and creating the evidence for it.

Everything is at traefik.io.