Les Jackson

Modern AI systems involve orchestrating multiple LLM calls, tool calling, human approvals, and long-running tasks. The challenge now isn’t getting an LLM to produce a coherent answer, it’s everything that happens around it. Failure is inevitable, especially in cloud native environments where services timeout, containers restart, pods are rescheduled, and external dependencies become unavailable. So how do we build AI systems that continue making progress when failure is simply part of normal operation? In this session we'll explore why reliability has become one of the defining challenges of production AI systems, and why traditional request-response programming struggles once work spans minutes, hours, or even days. Using practical examples, we'll examine architectural patterns for building AI systems that can survive failure, recover automatically, coordinate long-running work, and continue making progress without sacrificing the elasticity and resilience expected of cloud native applications. The focus isn't on prompt engineering or choosing the latest model. It's on designing AI systems that behave predictably in production, where failures are inevitable and reliability matters just as much as intelligence.