The gap between a working LLM demo and a production AI system is wider than most teams expect. A single prompt call is straightforward - but what happens when you need multiple models collaborating, tool use mid-conversation, human approval gates, and graceful failure handling at scale?

This talk dissects the architecture of real AI agent systems built for commercial clients in the media industry - including work analysing what the UK watches on YouTube, building on experience with audience measurement at BARB and AI-powered content classification at scale for video advertising platforms.

Starting from a single prompt-response pair, I'll progressively build up through prompt chains, tool-augmented agents, and multi-agent orchestration - covering the engineering decisions at each layer. Expect practical topics that rarely make it into tutorials: context management, sequential versus parallel execution, human-in-the-loop design, and why observability matters more than model selection.

Drawing on production systems built with Python, TypeScript, and modern LLM tooling, I'll share patterns that worked, patterns that failed, and the trade-offs that only become visible once real users and real data are involved.