2026-03-04
Governing AI Agent Operations in Kubernetes
Intended Team · Founding Team
Status: in development — admission controller real, distribution on the roadmap. The Intended Kubernetes admission controller exists in code and is proven in CI (helm lint+kindapply + a Calico egress-deny enforcement test). What is not yet published is the public Helm chart and container images — there is nocharts.intended.sorepo and noadmission-controllerrelease tohelm installtoday. We removed the earlier install snippet because it referenced an unpublished chart. Track this in the capability build roadmap and the Capability Truth Matrix (row: *Kubernetes / Helm → Partial*).
Where AI Agents Run
Kubernetes has become the default runtime for AI agent workloads. Agent frameworks deploy as pods. Tool servers run as services. Orchestration layers use the Kubernetes API for scheduling, scaling, and service discovery. When AI agents operate in production, they almost always operate in Kubernetes.
Why RBAC isn't enough
Kubernetes RBAC answers "which service account may call which API." It does not answer the questions AI-agent governance needs: *what is this agent actually trying to do, how risky is it, does policy allow it right now, and is there cryptographic proof of the decision?* RBAC is coarse and static; agent intent is fine-grained and contextual. That gap is what an authority layer fills.
The admission-controller approach
The model is a Kubernetes admission controller that intercepts agent-driven operations and submits them to the Intended authority engine before they are admitted: classify the intent, score the risk, evaluate policy, and admit only on an ALLOW decision — with every decision written to the tamper-evident audit chain. The controller fails closed: if the authority engine can't be reached, the operation is denied, not waved through.
That controller is implemented and tested today (packages/k8s-admission). What's left to make it a one-command install is publishing — pushing signed images and a versioned Helm chart to a public registry — which is tracked on the roadmap above.
What you can do today
While the published chart is in progress, the same enforcement is available through the gateway and SDK paths that *are* shipped (the authority API and the framework adapters). When the chart and images publish, runnable helm install instructions will land here.