AgentTest: An Intelligent Platform for Testing and Evaluating AI Agents Using Execution-Level Analysis

Agents built on large language models (LLMs) are now trusted with tasks such as order handling, refunds and customer support because they can call tools and APIs on behalf of an organisation. Their behaviour, however, is difficult to predict. An agent may select the wrong tool, act on wrong data, break a business rule, hallucinate, or claim success after a failed operation, while its final reply still reads convincingly. Testing only the reply therefore leaves such faults hidden, and existing benchmarks mostly measure general capability rather than the correctness of a particular agent inside a particular business workflow. This paper presents AgentTest, a platform that executes an agent against realistic scenarios, records its complete execution (input, tool calls, tool results and final response), and judges the run on four criteria: task success, action correctness, policy and safety compliance, and final-response accuracy. Failed runs are explained through a likely-cause analysis, an execution timeline and a consolidated report. The platform is designed with React.js, Node.js, Express.js, MongoDB Atlas and the OpenAI/Gemini APIs, and is illustrated on order-management and refund agents. The proposed approach gives developers a dependable way to verify agents before deployment and forms a base for automatic test generation, regression testing and production monitoring.