Abstract: Automatically grading handwritten answers to subjective questions in math exams is one of the important application scenarios of AI-empowered education. The existing methods, which rely on multimodal large models to grade handwritten answers to open-ended questions using prompts, face bottlenecks in accurately recognizing multi-line answers and ensuring fairness due to opaque grading criteria. This study proposes a multi-agent collaborative intelligent processing framework that divides the automatic grading task into subtasks and assigns them to four types of agents: the supervisor agent responsible for cross-agent collaborative coordination; the recognition agent responsible for handwriting text recognition; the analysis agent responsible for generating grading benchmark images and secondary analysis; and the grading agent responsible for matching scoring points. The specialized and complementary collaboration among agents not only breaks through the bottlenecks of single multimodal large models in automatic grading tasks but also provides scalability for future grading tasks with higher difficulties through modular design. The experiments in this study use basic math open-ended questions and their handwritten answers as test samples. Comparative experiments were conducted with GPT-4o and Qwen-VL-30B for handwriting recognition, and with DeepSeek-V3.2, Doubao-Seed-1.6, and Qwen3-Plus for human-machine grading comparison. The experimental results show that the overall average edit distance and accurate recognition rate in handwriting recognition are 0.40 and 85%, respectively; in scoring, the average human-machine absolute error, root mean square error, and consistency rate are 0.96, 1.92, and 61%, respectively, all outperforming single large model methods. This provides a feasible technical reference for the implementation of efficient and transparent automatic grading systems in intelligent education.