When AI agents fail, it's rarely the model's fault — it's because they weren't designed, they were just prompted into existence. Teams treat agent creation as prompt engineering when it should be software engineering. The result? Agents that work inconsistently, can't be maintained, and break when requirements change.

I'll present a framework for designing agents as engineered products with clear architecture, testable behavior, and maintainable structure. You'll see a production agent decomposed component by component, walk through real agents that fill roles across the software development lifecycle, and leave with a repeatable process you can apply immediately.