Everyone can build an agent that dazzles in a demo. Almost no one talks about what happens at 3am when it loops forever, burns £400 in tokens, or confidently calls the wrong tool against production data. This session is about the gap between the two.

Drawing on real deployments of agentic systems at Intent HQ, I'll walk through the architecture that actually holds up: how to distinguish a genuine agent from a workflow or a chain, and the core patterns tool use, memory, planning, multi-agent collaboration, and human in the loop, that determine whether your system reasons or just rambles. Then we'll get into the parts the demos skip: the failure modes (hallucinations, runaway loops, cost spirals, unreliable tool calls) and the production scaffolding that contains them, observability, guardrails, evaluation frameworks, and security. You'll leave with both the conceptual foundations and the concrete patterns to build agents that deliver value outside a controlled environment, plus an honest map of where agents genuinely excel today and where they still fall over.