On the Nature of AI Agent Failures on DeepSeek V4 Pro and Gemini 3.8 Flash
Year: 2026
Type: article
Abstract: This document presents an analytical breakdown of the behavior of AI agents that call functions with real effects. The agents operated on the DeepSeek V4 Pro and Gemini 3.8 Flash models. The "Warehouse" agent testing system is built on the methodology whose main provisions we set out previously (doi.org, habr.com). Subsequent innovations concerned primarily the optimization of the method, the expansion of the testing system’s analytical capabilities, and the improvement of the interface and visualization of conclusions in run reports, while the main theoretical provisions remained unchanged. To verify whether an AI agent will call a function that performs an action with a real effect when it must not be called because the unstated conditions of its application are not met, and to compare the action performed with the response the agent gives to the user. The conclusions concern the character of failures in upholding unstated conditions. In particular, it is shown that DeepSeek works with a condition as an object of reasoning, whereas Gemini treats it as a signal; an analysis is provided of which of the identified layers (communicative, logical, or cognitive) the failures occur in for each model; and the agent’s reporting to the user as well as its responses to a persistent remark from the operator are examined.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Keywords: function calling, DeepSeek-V4-Pro, LLM agents
Open Access status: green
OpenAlex ID: W7225919994
