Back to Schedule

Claw Native: Building a Kubernetes Operator for Multi-Agent AI Workloads

Alessandro Stefouli-Vozza

10:15 - 11:00

Pavillon

Most "agents on Kubernetes" stories end at a Helm chart and a Deployment. That works until you have fifteen agents across four channels, each with its own LLM provider keys, rate limits, channel webhooks, and lifecycle that doesn't fit a Deployment's worldview. At that point you either bolt more YAML onto the side, or you make agents a first-class Kubernetes resource. This talk walks through the design and implementation of an OpenClaw operator: CRDs (Agent, AgentChannel, LLMProvider), a controller built with controller-runtime, and the reconciliation patterns that actually survive production. We'll cover the API design decisions that matter — why Agent is not a Deployment even though it ships pods, how to model channels (Discord, Slack, Telegram, WhatsApp) as their own reconciled resources, and how LLM provider credentials, virtual keys, and fallback routing become declarative through a LiteLLM-aware controller. Then the hard parts: status subresource design when your workload's state is "talking to a user for the next 40 minutes," finalizer ordering when an agent run is mid-flight, secret rotation without dropping in-flight conversations, and the operational anti-patterns that look right in kubebuilder init and bite you at the third on-call. You'll leave with a concrete reference for designing operators in the agentic-AI space, the four reconciliation patterns we landed on after rejecting six, and an honest read on when extending Kubernetes is the right call versus when you're just building a worse version of a Helm chart.