AI systems don’t fail cleanly. They drift, degrade, and behave just well enough to pass… until they don’t. What looks stable in testing starts to shift under real usage. Context gets messy, tool calls fail in subtle ways, and outputs vary just enough to create risk. It’s not just what the model does, it’s everything it depends on. This keynote unpacks what building AI actually means once you move past the prototype, where things start to break, why they’re hard to catch, and how teams are adapting in practice. Key takeaways: → Where non-determinism turns from a quirk into a reliability problem → Why context, tools, and orchestration introduce hidden failure points → How teams are adapting when systems don’t behave the same way twice

Agent systems are hitting real production limits - unreliable tool calls, broken workflows, context loss, unpredictable decision-making, and systems that become difficult to debug once autonomy increases. As enterprises push agents beyond prototypes and into live environments, many are discovering that orchestration, observability, permissions, and reliability become significantly harder at scale. This panel explores the gap between agent design and production reality, unpacking the operational and architectural challenges emerging as autonomous systems move into enterprise deployment.