Traditional software executes predefined logic in response to requests. Agentic systems observe their environment, make decisions, and act over time. This session introduces the core distinction that defines agentic AI and explains why existing mental models for building and operating software no longer fully apply. We’ll examine how state, memory, and decision authority change system behavior, introduce new failure modes, and complicate responsibility in production environments. The focus is on understanding what engineers must reason about differently once systems are allowed to act autonomously. This session sets the foundation for discussions on coordination, evaluation, and control throughout the Agentic AI track.

Generative and agentic AI behave differently once embedded in real systems. This session examines what changes when AI moves beyond experimentation and begins influencing workflows, decisions, and execution. The discussion covers common failure modes in operational AI systems, the risks introduced by agentic behavior, and the design considerations required for control, observability, and recovery. It also looks at how to integrate AI into existing systems without disrupting delivery or creating long-term technical debt. Key takeaways: → How and where generative and agentic AI systems fail in real environments → The risks agentic behavior introduces and how to design for oversight → Practical patterns for control, monitoring, and recovery