Tool-using agents can fail twice: the tool fails, and then the agent tells the user it succeeded anyway. The authors introduce Failure-Transparent Agents (FTA), a 100-task that fixes the failed tool observation before generation, so post-failure claims can be audited against evidence the model actually received. Across six models and 3,600 human-annotated responses, false-success rates were 22.8% with a baseline prompt, 9.3% with a transparency instruction, and 0.8% with a structured evidence contract.
Fabricated detail rates followed the same pattern, dropping from 28.3% to 14.3% to 0.8%. Useful responses increased from 74.9% to 89.2% to 98.8%. The finding is interesting because it isolates reporting failure from tool selection and recovery, and shows that a simple output structure can dramatically reduce unsupported claims without making agents useless.