AI agents are moving from impressive demos to systems that can read company data, call APIs, execute code, modify infrastructure, and trigger real business processes. That changes the engineering problem completely. A traditional application follows paths we designed in advance. An agent decides what to do at runtime. Once we give it tools, credentials, memory, and access to production systems, we need to answer a much harder question: how do we keep an autonomous system useful without giving it unlimited trust? In this session, we will build a practical production architecture for agentic AI around four control boundaries: identity, policy, observability, and containment. We will explore how to: give agents their own identities instead of borrowing user credentials; apply least-privilege access to tools, APIs, and data; control what enters and leaves an agent's context; observe prompts, tool calls, decisions, latency, failures, and cost; introduce approval gates for high-risk actions; detect unexpected behavior; design rollback, timeout, quota, and kill-switch mechanisms; operate agents under enterprise security and governance requirements. The focus is not on another chatbot demo. It is on the engineering decisions that appear after the prototype works and someone asks: “Can we safely put this in production?” Attendees will leave with a practical reference architecture and a checklist they can apply to their own agentic systems, regardless of the model or cloud platform they use.