Google engineers have put together a 90-minute workshop showing how AI development is shifting away from basic RAG (Retrieval-Augmented Generation) toward more sophisticated graph-based approaches. The progression they outline moves through RAG, then Graph RAG with semantic understanding, into memory systems, and finally multimodal agent graphs that can orchestrate specialized agents working together. This isn't theoretical—they're walking through actual production setups, not toy examples. The workshop breaks down into specific segments: the first 15 minutes cover setting up a production agent stack from scratch, then around the 30-minute mark they dig into Graph RAG with semantic capabilities, which is where things get interesting. By the 65-minute point they're showing how to extract useful context from graphs, and the final stretch focuses on orchestrating multiple specialized agents to work in concert. Each section builds on the previous one, so you're watching the actual progression from simpler retrieval methods to more complex reasoning systems. The core shift here matters because basic RAG has real limitations—it just finds relevant documents and feeds them to an LLM. Graph-based approaches create structured relationships between concepts, giving the AI better context and reasoning ability. When you add memory and multiple agents that can handle different tasks, you get systems that can actually handle complex problems instead of just pattern-matching against stored documents. The workshop apparently makes this transition concrete enough that you can actually understand what's happening at each stage.