Everyone's building AI agents now, but there's a massive gap between prototype and production. I learned this the hard way whilst building SheerFit, an AI fitness coach that's been running in production for over a year. This talk covers three things that actually matter when you're trying to deploy agents: **Model Context Protocol (MCP)**: The new standard for connecting agents to tools and data sources. I'll show you how to build MCP servers in Python using FastMCP, and why this matters more than whatever custom integration you're currently maintaining. MCP was only released in November 2024, so this is pretty cutting-edge stuff. **Persistent Memory**: Agents without memory are just fancy chatbots. I'll walk through how we implemented semantic memory in SheerFit using Redis and sentence transformers. This lets agents remember context across conversations whilst keeping retrieval under 500ms. I'll show you the actual architecture and where we got it wrong the first time. **Multi-Agent Orchestration**: When do you use LangGraph vs CrewAI vs AutoGen? I've tried all three. I'll compare them honestly, show you which patterns work (and which ones will have you debugging conversation loops at 2am), and demonstrate how to coordinate multiple specialist agents without things falling apart. Throughout the talk, I'll focus on self-hosted, open-source approaches. Not because I'm ideologically opposed to cloud services, but because they're expensive and you don't always need them. The patterns I'll show work on modest hardware and won't leave you with a massive AWS bill. You'll leave with actual code you can use, honest comparisons of the major frameworks, and production deployment patterns from systems that serve real users. No hand-waving, no "this should work in theory" just stuff that actually works.