What if you could have a whole “senior engineering team” review your code—locally—using open-source LLaMA 3? In this video, I vibe-code a real AI Coding Reviewer Agent system using the AutoGen framework and a locally running LLaMA 3 model via Ollama. No OpenAI, no cloud APIs—just a multi-agent setup that collaborates like humans: back-and-forth discussion, critique, iteration, and validation. I first zoom out and break down what AutoGen actually is: not prompt-chaining, but conversation orchestration. An agent is basically an LLM + instructions + behavior rules (system messages), and the magic happens when you put multiple agents into a GroupChat with a GroupChatManager that controls turns, context flow, and stop conditions. Then we build a three-agent “scrum team”: a grumpy Senior Code Reviewer, a Refactor/Code Correction Agent, and a Test Case Generator—plus a User Proxy Agent to kick things off without a human-in-the-loop. Along the way, I also talk about why AutoGen feels closer to real collaboration than typical pipelines, and I’m honest about the struggle: version confusion, compatibility issues, and LLaMA 3/Ollama timeouts. The key takeaway: agentic systems are designed like teams—roles, prompts, and orchestration matter as much as the model.