What happens when the agent makes a mistake?

We plan for it from the start, because an agent usually does not crash – it returns a plausible-looking piece of nonsense. That is why every step has a check against hard data: verification in the catalogue, a control calculation, a comparison with the database. Whatever fails the check goes to a person, together with the reason it was stopped. Every tool call is logged including inputs and results, so for every output you can trace back what it was based on.