An answer needs to carry its origin
A reliable agent separates sourced information, inference and what could not be determined. Its knowledge base should record version, owner and validity. When information changes, the team needs to know which answers and decisions may have been affected.
Boundaries are part of the interface
Users should understand when an agent may act, when it only recommends and when it hands the task to a person. Escalation is not a technical footnote. It is a core product function, especially in health, finance, contracts and irreversible decisions.
Test behaviour, cost and recovery
Evaluation should include normal cases, ambiguity, conflicting instructions, missing data and tool failures. Logs should record model, prompt version, tools called, time, cost and outcome without exposing secrets. Shutdown criteria complete the design: a responsible agent knows when to stop.