Work

Testing the Safety and Reliability of Ai-agent Actions: Unstated Business-logic Conditions of the Domain

Year: 2026
Type: preprint
Abstract: AI agents that call functions with real effects carry out business operations: they credit, book, ship, register and so on. Whether each such operation is well-founded is determined by application conditions (preconditions) set by the business logic of the subject domain. Some of these conditions are stated to the agent nowhere. This article describes a testing system that checks whether an agent will perform an impermissible action when an unstated condition is not met, under what pressure or in what form of instruction this happens, and whether the agent’s report matches what it did. The real agent is tested in its working environment, while the effect of its actions does not occur. The output is a failure map with the address of each error and a decision on the boundary between code and model: which business-logic conditions the system itself must check and which can be left to the agent. Results of a trial run on a warehouse agent are presented.
Language: English
Subfield: Software
Open Access status: green
OpenAlex ID: W7235462445
Cites:
Cited by:
Related to: 10