RESEARCHInvestigateNOW
Fabrication After Tool Failure: Tool-Augmented Agents Assert Values Their Tools Did Not Return
arXiv cs.CL — Computation and Language
Factual evidence
What the source reports
Benchmark study finds tool-augmented AI agents fabricate values 14.1% of the time when external tool execution fails.
OneBench interpretation
Institutional assessment
So what
Silent agent hallucinations following API failure expose automated financial workflows to unvalidated data insertion without throwing runtime errors.
Do what
Review error-handling and payload validation controls with the team responsible for agentic workflow architecture.