Frontend Masters reposted this
"RAG is not everything. There is too much marketing on that." That line hit me during a two-day AI Engineering workshop at Frontend Masters HQ in Minneapolis, taught by Scott M. We built an AI agent from scratch across twelve lessons. Real code, real evals, and real tool design. By the end, the agent could generate and modify technical diagrams on a canvas, search a private knowledge base, and self-correct based on automated feedback. The biggest lesson wasn't any single technique. It was the following loop: Build something -> Measure it with evals -> Find where the eval is lying (it usually is) -> Fix the measurement -> Fix the agent. One change at a time. Three iterations out of five were about fixing the eval, not the agent. That surprised me. The instinct is always to change the code. The discipline is to first make sure you're measuring honestly. Other things that stuck: - Everything is context. The system prompt, the canvas state, the tool descriptions, and the retrieved documents. It's all tokens going into a context window. Getting the right tokens in the right shape at the right time is the actual job. - Workflows for known repeated paths are where the money is. If you know the sequence, don't make the model figure it out every time. - The data flywheel is the first thing to build. Log every interaction, add a thumbs up or down, and promote failures into your eval dataset. Start on day one. I flew in from Brazil. It was like -7°C. The coldest temperature I've ever experienced. Worth every frozen minute. Now applying this at work and in side projects.