Abstract
Artificial Intelligence (AI) agents are now capable of performing complex reasoning, tool usage, research assistance and problem-solving activities. However, there is currently no standardized and automated mechanism to evaluate the scientific reliability and research quality of such AI systems. As a result, many AI agents produce outputs that appear logically correct but are not verified or scientifically valid. This creates a major risk in domains where accuracy and trust are critical. To address this challenge, we propose an AI Agent Testing and Analysis Platform: a unified automated evaluation system that benchmarks AI agents across multiple research stages. The platform executes research tasks, captures agent responses, and evaluates them based on multi-dimensional quality metrics such as relevance, completeness, clarity, structure, depth, accuracy, creativity and coherence. The evaluation process is fully automated using the concept of LLMas a-Judge, where advanced LLMs score the responses using structured scoring rubrics. Experimental results show that the system can consistently detect strengths, weaknesses and failure patterns across agents, while maintaining strong alignment with expert judgment. This work establishes a scalable, framework independent and repeatable foundation for benchmarking AI research agents, enabling transparent comparison, objective scoring and improved scientific trustworthiness.