The ToolReplay tool has been released for auditing calls of AI-agents tools

Developers presented ToolReplay — a specialized tool for analyzing logs of AI agents' work. The solution is designed to check transcripts of tool calls for errors and vulnerabilities. The system performs an audit in three key areas: searching for non-deterministic behavior, identifying redundant calls, and controlling access rights exceeding. The tool is written in Python 3.11+ and does not require the installation of external dependencies, which simplifies integration into existing development pipelines. The project is focused on developers creating autonomous systems, where it is critically important to track the correctness of the agent's interaction with external APIs and functions. Using ToolReplay allows automating the debugging process of complex action chains that the model performs during task execution. The tool is available as an open repository on GitHub. The absence of external dependencies makes it a lightweight solution for local log analysis without the need to send data to external servers, which increases security when working with sensitive agent data. The introduction of such audit systems is becoming a standard in the development of AI agents, as it allows minimizing the risks of "hallucinations" when calling functions and preventing unauthorized actions of the model in the target environment.