Do AI Agents Need a Semantic Layer?

Everyone in data has an opinion about semantic layers right now. Half the industry says AI agents are useless without one. The other half is shipping agents that work fine on a pile of markdown files. Jacob Matson and Alex Monahan have been testing the question properly: same warehouse, same questions, one agent grounded in plain-text context and one in a real semantic layer built with Malloy. They'll share what the evals actually showed, where plain context holds up, and where determinism starts to earn its cost. The answer even changes as model prices drop.