An agent can complete the task and still fail the business case. Every extra step has a cost: longer context windows, more model calls, tool use, retries, latency, and infrastructure overhead. What looks impressive in testing can quickly become too slow, too expensive, or too unpredictable to run at scale. This session explores where cost starts to compound in agentic systems, why optimisation has to be designed in early, and how teams are making trade-offs between autonomy, performance, and affordability without stripping away the value of the agent. Key takeaways: → Where cost builds up across agentic workflows → Why latency, retries, tools, and context windows change the economics of AI systems → How teams are designing agents that are useful, reliable, and affordable to run at scale