A generative output that a person reads is easy to forgive. One that feeds the next step in a workflow is not. Most generative stacks were designed for one-off results: a prompt, a retrieval step, and an answer someone checks. Once those outputs start feeding other steps and triggering real actions, the prompting, retrieval and evaluation choices that worked in isolation start to break. So what has to change in the stack before it can carry real work? This session brings together practitioners who are rebuilding their generative stack around workflows. They'll cover where their original design stopped holding up and what they changed. Through real examples and open disagreement, the panel will get into how prompting, retrieval and evaluation shift when outputs become dependencies, and how they decide when a system is ready to act without a person checking every step. Key takeaways: → Where stacks built for one-off results break once outputs feed other steps → How to design prompting, retrieval and evaluation for outputs that trigger actions → What to put in place before letting a system act on its own The questions still line up with the new takeaways: