We ran into similar issues when experimenting with multi-step agents. The biggest gains didn’t come from “smarter” prompts, but from reducing how much freedom an agent had at each step.

What helped for us was forcing explicit boundaries: what context is allowed in, which tools are visible, and what output is valid before moving forward. Once those constraints reset between steps, reliability went way up and hallucinations dropped sharply.

I’ve been using GTWY for this kind of step-gated setup and it made these boundaries much easier to reason about. The agent feels less clever, but far more predictable, which is what you want in real workflows.

Curious how you think about resetting authority between steps vs letting context accumulate.