In this video, I vibe code a Productivity AI Agent from a completely blank folder in VS Code. The goal is simple: take a meeting transcript as input, then summarize it, extract action items per attendee, and propose follow-up meetings with the right participants. I build it in Python using LangChain, a ReAct Zero-Shot Agent (so the LLM can decide when to use tools), and a locally running LLaMA 3 model via Ollama—no UI, just clean console output so you can see the agent working end-to-end. I also show how I use GitHub Copilot Agent Mode with Claude Sonnet 4 for code generation, but I’m very deliberate about the prompt. I write a requirements file that defines outcomes, constraints, measurable success criteria, and even pin a Python version—because dependency mismatch is one of the fastest ways to waste time when you’re prototyping. After generating a realistic transcript, I review the generated code, walk through the core agent setup (tools, prompt template, agent executor), then run it locally and validate the outputs. Finally, I talk about what to improve next: a richer UI with transcript upload, faster hosted LLM calls for production, proper tests, and observability. And the fun part—integrating tool calls with real systems like Jira/Trello for action items and Google Meet/Outlook for scheduling so it actually behaves like a real assistant.