A multi-step research agent compiles supplier risk reports. Reviewers notice that when one data-source tool fails, the final report still states "no adverse findings" for that supplier with high confidence. The agent's self-reported confidence field is consistently above 0.9. What is the root cause to address?

Answer: B — The tool returns access failures in a form indistinguishable from a successful search with no results

The agent can't tell "couldn't check" from "checked and found nothing," so a tool outage turns into a false all-clear. Structured errors that distinguish access failures from valid empty results, plus coverage annotations, fix the cause. (a) relies on uncalibrated self-reported confidence, which here is high even when the report is wrong. (c) and (d) don't give the model the information it's missing.