AgentTester: Improve Coverage in Automated Test Generation With Multi‐Agent as Optimizer

Heuristic algorithms have been widely applied in search‐based software testing (SBST), playing a critical role in the automated generation of high‐coverage test cases. These algorithms often rely on random operations such as initialization, crossover and mutation to provide diverse individuals for search. However, it is possible to cause low‐quality solutions and a waste of resources due to their undirected randomness, which can be replaced by clear end‐to‐end reasoning. For large language models (LLMs) have demonstrated exceptional capabilities in handling simple reasoning tasks with the Chain‐of‐Thought (CoT) scheme, this paper introduces AgentTester, a multi‐agent framework of LLM, where subtasks in the search process are assigned to multiple agents applying CoT to efficiently generate high‐quality test cases. Experimental results show that AgentTester outperforms comparison baselines across benchmarks in generation with high executable rate, coverage, readability and convergence speed. Results validate the effectiveness of each agent and further underscore the potential of LLM in addressing complex reasoning tasks.