AgentBench is an open-source benchmark repository containing 138 real-world Python programming tasks for systematically evaluating the performance and cost of large-language-model coding agents. It is intended for AI researchers and developers who need a standardized suite to measure agent effectiveness under different repository context conditions.