Why do capable AI agents lose context, repeat mistakes, and declare victory too soon? Drawing on research and lessons learned building harnesses used by tens of thousands of enterprise users, this talk explores how better knowledge management frameworks, deliberate context management, durable memory, and verifiable feedback turn fragile demos into dependable workflows. It examines how agents can preserve the right information across tasks, conversations, and tool calls without becoming overloaded by irrelevant history. It distills practical lessons that open-source developers can apply immediately to build agents that are more reliable, observable, and effective in real-world environments.