Agents fail in boring ways. A tool times out, a model call errors, a worker restarts, or a human approval takes six hours, and the agent loses its place.

Cadence is a CNCF sandbox project for durable workflow orchestration, and the engine Uber runs its production agentic workflows on. In five minutes, I will show what changes when an agent runs as a Cadence workflow: state is persisted, failed steps retry, approvals can wait as long as they need, and every decision lands in a history you can inspect.

You will leave with a simple test for when an agent should become a durable workflow, a clear picture of where Cadence sits in the AI stack, and a way in: docs, samples, and an invitation to our ContribFest.