KubeCon EU 2026: MCP Becomes an Enterprise Protocol

Back from KubeCon EU 2026 in Amsterdam. AI agents are becoming part of the cloud native baseline, not an add-on. The talks I kept coming back to sat at the overlap of Kubernetes, platform engineering and AI infrastructure.

MCP Becomes an Enterprise Protocol

David Soria Parra from Anthropic opened Agentics Day with "MCP in 2026: Context is All You Need" — and it resonated because it spoke to what every enterprise in the room was dealing with: connecting agents to systems at scale, across teams and security boundaries.

His argument: 2026 is the year MCP becomes an enterprise protocol — standardized auth, governance, remote connectivity, introspectable by design. The industry is converging, and an open standard matters more than waiting for a perfect one.

The Coordination Layer for AI Traffic

The other thing everyone kept circling back to was what sits in front of all this: a coordination layer for AI traffic. The challenge is different from traditional API gateways because agents make autonomous decisions at runtime about which tools to call and which data to access. You cannot govern that with static routing rules alone.

As MCP servers multiply, enterprises need a thin layer that enforces policy, identity and data boundaries dynamically, across many backends, without adding another approval queue. The moment this layer creates friction, teams will route around it. It has to simplify, not block.

We have been working through similar questions at E.ON with our Agent Gateway work — trying to build a lean layer where security and compliance are built in and the approved path is faster than the shadow path. KubeCon confirmed we are asking the right questions.

Readiness Criteria, Not Demos

Agentics Day shifted the conversation from impressive demos toward readiness criteria for platform and compliance teams: defined error budgets for hallucinations, end-to-end traceability — with OpenTelemetry emerging as the way to harmonize agent observability as it did for microservices — and runtime enforcement instead of design-time reviews only.

The bar everyone kept coming back to: the human effort to supervise agents must be lower than the work they take off the table. If it is not, you do not have an agent. You have a liability.

Context as a Versioned Artifact

The other shift I noticed was in how teams talk about context. Skills, knowledge files and configuration are moving from ad hoc prompt snippets to versioned artefacts with tests, evaluation pipelines and drift checks — much closer to how infrastructure is managed today. That is the real maturity signal.

The most useful questions leaving KubeCon are about architectural fit: what is the strategy for MCP-style integrations, where do gateways sit in the landscape, who owns the lifecycle of context and skills, and critically — is your governance layer something teams want to use, or something they work around? Those questions will shape whether AI agents become a natural part of cloud native environments or stay bolted on at the edges.