Build long-running agents with Google’s Agentic Stack | The Agent Factory

In this episode, we explore "long-running agents": AI systems that persist, sleep, self-correct, and execute complex workflows over hours, days, or even weeks. Join host Smitha Kolan and Addy Osmani, Director of Google Cloud AI, as they break down the three non-negotiable rules for production-grade agents: durable state checkpoints, event-driven dormancy (true sleep), and separated evaluation. You'll see three live demos of long-running agents: a multi-day HR onboarding coordinator that delegates tasks and waits for human signatures; a custom operating system built autonomously; and a complex, highly optimized 3D video store generated entirely in Blender by an AI agent over several days. Whether you're a cloud architect, an AI engineer, or a technical founder, you'll walk away with a blueprint for building long-running agents that actually finish the job. You'll learn how to implement the three-agent setup (planner, generator, evaluator).