The Agent Harness Control Plane: Running AI Workloads Safely on Kubernetes

What happens when your on-call pager goes off because an AI agent entered an infinite loop and burned $4,000 in API tokens overnight? As enterprise teams push autonomous agent swarms and Model Context Protocol (MCP) implementations into production, they face a brutal realization. AI agents are only as reliable as the underlying infrastructure. When a tool call drops, a container restart kills a long-running task, or image-pull throttling stops a workflow, it is not an AI problem. It is a cloud-native infrastructure failure. In this lightning keynote, we shift the conversation from prompt engineering to platform engineering. We will break down how to build secure, deterministic infrastructure for non-deterministic AI workloads. Discover how to leverage Kubernetes Dynamic Resource Allocation (DRA) to eliminate GPU cold starts, isolate untrusted agent code using Docker Sandboxes, and trace tool execution safely. Your agents are a distributed system. It is time to run them like one.