Can you teach an AI to follow your workflow instead of managing it constantly?

Current AI agent systems use external orchestrators that repeatedly prompt a frontier model—expensive and context-heavy. Instead, the authors embedded complete workflows (like a 55-node insurance claims process) directly into small fine-tuned models' weights. Results: near-frontier quality on travel booking, customer support, and claims processing at a fraction of the cost, with no need to expose procedures to third-party APIs. Code and models are released.