NEW YORK — What happened: A comprehensive audit of enterprise-grade function-calling models analyzed how frontier LLMs handle complex API integration pipelines, parameter mapping, and unexpected JSON responses.
Why it matters: Autonomous AI agents rely entirely on accurate tool use. If a model passes incorrect parameters or misinterprets an error code from an external database, it can trigger cascading failures across automated business workflows.
Deep dive: The audit tested models against dynamic API schemas that shifted slightly mid-task. While models excel at static function calling, they frequently exhibit 'silent failures'—where they fabricate plausible-looking default arguments instead of halting and asking the user for clarification when required parameters are missing.
Report check (claims vs what is verified vs still rumor): Providers claim near 99% accuracy in function-calling benchmarks. However, independent testing verifies that accuracy drops below 60% once APIs require stateful multi-turn interactions or return nested error objects.
Open questions: Can self-correcting critique loops reliably catch these silent argument fabrication errors before they impact production databases?