Agent Substrate, with Tim Hockin and Brandon Royal

Agent Substrate introduces a new set of primitives designed to address the unique challenges of running AI agents at scale on Kubernetes. Unlike traditional web services, agentic workloads are highly non-deterministic and exhibit significant idleness, often exceeding 90% while waiting for user input or model responses. By decoupling the agent lifecycle from the underlying infrastructure through an actor-worker architecture, Agent Substrate enables efficient multiplexing and rapid startup times. Key features include native suspend and resume capabilities, which allow state snapshots to be stored in durable storage, effectively freeing up compute resources during idle periods. This approach leverages existing Kubernetes ecosystem tools for scaling and management while bypassing control plane limitations to support massive agent density. Tim Hockin and Brandon Royal highlight this as a necessary evolution to accommodate the shift from static web applications to dynamic, always-on agentic systems.