This matches our experience too. The biggest reduction in hallucinations usually comes from shrinking the action space, not improving the prompt. When inputs, tools, and outputs are explicitly constrained, the model stops “being creative” in places where creativity is actually risk.

It’s less about smarter models and more about making the system boring and deterministic at each step.