Testing AI Agents That Call Functions with Real

Abstract

This document sets out a methodology and a method; it contains no empirical results. The methodology is being tested under a separate plan; the course of testing will be reflected in new versions of this document, and the results and their analysis are planned for separate publications. Aim. To detect errors of an AI agent that calls functions with real effects: a function call made when the function’s application condition is not met, and a mismatch between the agent’s message and the action it performed. For each error we state its type — which condition was violated — and its character: under which instruction format it occurs and how it is reflected in the agent’s message. The result serves to define the boundary between code and model, that is, to decide which condition checks are performed by the function’s code and which are left to the agent, and to monitor this boundary when the agent’s model, prompt or functions change.