Beyond the Prompt: Building AI Harnesses That Actually Work

Everyone's building AI agents, and most of them break in ways nobody can explain. The model usually gets blamed, but the real problem is the harness: the routing, context handling, and debugging infrastructure around it. In this session, Gezim Haxhia (founder, dodorouter.com) breaks down how to design AI harnesses that actually work. You'll learn how to debug agents when nothing "crashed" but the output is wrong, and how to make your AI stack's failures visible instead of mysterious. No deep ML background needed.

Gezim Hoxha is the founder of DodoRouter.com, a model-routing and debugging platform for AI applications. A serial entrepreneur, he previously built ImagineAPI.dev, an unofficial Midjourney API that grew to thousands of developers before Midjourney sued him over it. He won his users' trust, lost the legal fight, and kept building. He now focuses on the unglamorous infrastructure that makes AI systems actually work: routing, harnesses, and observability.