Most agents demo well and fail in production. After seeing hundreds of agents up close and studying what actually ships, clear patterns emerge that separate agents that reliably get the job done from those that don't. This talk covers design patterns and techniques for production agents: everything from memory architectures that maintain coherence, tool design that outweighs model selection, human-in-the-loop checkpoints, verification loops that catch errors long before users do, graceful failure recovery, and evaluation strategies when ground truth doesn't exist. Practical, implementation-ready, no hype