Uber's AI/ML platform, Michelangelo, runs its agents on Cadence. Its Agent Executor Service uses Cadence as the durability layer for agentic workflows in production, serving teams across the company. An agent that crashes, stalls, or waits days on a human picks up exactly where it left off.

This session explains why that layer is needed and how to build it. Most agent SDKs are stateless by design, which makes them fragile where production is hardest: process crashes, hanging tools, and tasks that run for days. With Cadence underneath, a run resumes instead of restarting, every decision and tool call is recorded and replayable, and a long wait costs nothing.

We cover the architecture behind Uber's setup and what it takes to keep agents reliable at that scale. We then walk through the Cadence integrations for agent SDKs such as OpenAI Agents and Google ADK, and the AI coding tool skills that get you to a first durable agent workflow in minutes.

You will leave with a blueprint for making your agents as reliable as your microservices.