As teams move beyond simple “one LLM + prompt” prototypes, their stacks start to look more like systems: multiple models, agents, tools, data layers, and evaluation loops all stitched together. With that shift comes a new set of headaches unexpected behaviour at scale, fragile orchestration, unclear ownership, and architectures that are hard to evolve once they’re in production. In this session, engineering and product leaders unpack how they’re designing modular, multi-component AI systems that can still be understood, governed, and trusted. Expect candid conversations about when modularity actually helps, where it introduces new failure modes, and how teams are thinking about patterns like MCP, agent coordination, and shared infrastructure. Key takeaways: → How teams are structuring modular AI systems without creating brittle dependencies. → Architectural patterns that improve reliability as models, agents, and tools interact. → Where modularity introduces new risks—and how leaders are mitigating them. → How to design systems that stay adaptable as capabilities and requirements evolve