My recap of KubeCon + CloudNativeCon North America 2026 in Salt Lake City. The question all week was how to take agents to production. My takeaway: we already have most of the building blocks, but we need to rethink how we put them together, especially around governance, observability and autonomy.
A full day of MCP and agents at KubeCon + CloudNativeCon North America in Salt Lake City, which I co-chaired with Chris Mazur. Open governance, securing the tool layer and running agents on Kubernetes kept coming up, and most of it came back to platform engineering.
Three days in Malmö at Øredev's 20th edition, under the theme Legacy. Agents came up in almost every talk, but the conversations kept returning to older questions: where the bottleneck is, what the platform should carry, how much structure our systems have, and who we're building for.
My first KCD UK, and my first time in Edinburgh. Abby Bangser made the case that platform engineering is the prerequisite for getting value from AI, and the rest of the two days kept agreeing with her: agent sandboxes, security as a platform user, policy as the platform's API, and a Kubernetes API that survives whatever you swap out underneath it. Plus Hazel Weakly on strategy, Adriana's career lessons, and the talk Adriana and I gave.
My first Open Source Summit, and the end of a week in Prague. The idea that kept coming back was participation: in the governance of the projects Europe depends on, in defending maintainers now that AI makes vulnerabilities cheap to find, and in the instrumentation and guardrails that let us trust agents. Plus agentgateway, the Kubernetes Agent Sandbox, a benchmark of agent memory backends, and the talk Adriana and I gave.
The first Observability Summit Europe took place in Prague, five months after OpenTelemetry graduated. The talks I caught were about what happens after hello world: getting the Collector into production, making it stable by default, and getting GenAI telemetry to converge on OpenTelemetry.
Nearly every talk I saw at GOTO Copenhagen this year was about coding agents. The surprising part wasn't the AI. It was how often the answer came back to fundamentals we already knew mattered: code health, deterministic checks, clear boundaries, and platforms built for humans first.
Telemetry can already show what an AI agent did. Deciding whether it should have done it is the harder problem, and it came up on stage and at both dinners at SREday London.
goose speaks OTLP, so a local coding-agent session becomes a trace you can read: which tools ran, how long each model call took, and where the tokens actually went.
Most organizations have observability as a capability, not as a product. Platform engineering was Kubernetes' Rails moment; observability is still looking for its own.