Abstract: Leaf disease is one of the primary threats to apple yield and quality. However, existing visual algorithms cannot fully recognize disease types and local visual symptoms. Targeted disease treatment has then relied heavily on the human diagnosis from the experience of agricultural experts. It is still lacking in an end-to-end intelligent support system. Furthermore, current large reasoning models (LRMs) can be expected for autonomous agents. Severe limitations have often occurred, including single-path reasoning mechanisms, excessively high computational overhead, and overthinking to generate highly redundant and inefficient reasoning steps. In this study, an agent (apple leaf disease management agent, ALDMA) was developed using LRM with SPO-SDR (stepwise preference optimization-serialized decaying rewards) reasoning optimization. (1) In standard operating procedures (SOP), four expert systems were integrated into the agent: image preprocessing, disease detection, disease analysis, and treatment plan generation. Thereby, the agent was then integrated to bridge the operational gap between perception detection and executable treatment. (2) Candidate thoughts were generated for the datasets via a tree of thoughts (ToT) mechanism. Direct preference optimization (DPO) was employed to align reasoning paths with ToT preferences. (3) SDR mechanism utilized the rule and exponentially decaying rewards to regulate intermediate reasoning, thus forcing early generations of high-quality paths for early truncation. Experimental evaluations were conducted on the ALDOD dataset. The exceptional performance of the ALDMA framework was achieved in multiple dimensions. The YOLO11x model was selected as the optimal model at initial perception. A remarkable AP_0.5 of 98.4% and a mAP of 97.7% were obtained to maintain lightweight parameters. Three common diseases (rust, scab, and black rot) were selected in the generation tasks. ALDMA achieved outstanding resolved evaluation metrics of 86.21, 88.32, and 89.89 under 1.5, 7, and 14B parameter-scale LRMs, respectively. The average resolved score of 88.14 outperformed conventional and advanced baselines, including CoT (82.50), ToT (81.86), SFT (84.47), FPO (81.47), GoT (86.77), and S-GRPO (85.61). SPO-SDR integration was established for an optimal equilibrium between reasoning performance and computational efficiency. Multi-path exploration was shifted into the DPO training. ALDMA also maintained CoT-level efficiency during inference. Tasks were measured 10.89 times faster than that of the ToT approach. Token consumption and overall reasoning time were significantly reduced by an average of 30.9% and 31.2%, respectively, compared with baseline models. Ablation studies validated that SPO alone in a 7B model increased the Resolved metric by 6.2% to 88.42, while SDR alone reduced the token usage by 5.3% and execution time by 8.1%, in terms of the individual components. The combined SPO-SDR strategy on the 7B model achieved the ultimate optimal balance, thereby yielding a Resolved score of 88.32 with minimum token usage (4201) and time (185.40s). Moreover, cross-scale comparisons revealed that while the 14B model provided the highest absolute accuracy, the 7B model delivered the most practical trade-off for agricultural edge deployment. Furthermore, loss curve analysis verified that SPO converged to lower loss intervals with higher alignment to human expert preferences, compared with FPO. A seamless end-to-end workflow was established for the visual detection of apple leaf diseases to autonomously generate complex and executable plans. According to Standard Operating Procedure (SOP), four expert roles were seamlessly coordinated—image preprocessing, disease detection, disease analysis, and treatment plan generation—to remove the conventional reliance on human intervention. LRM can be expected to optimize ToT-level high-fidelity performance with CoT-level low-latency computational efficiency. The framework can provide a robust, highly efficient, and deployable intelligent solution to rectify the disconnect between agricultural diagnosis and practical governance. The findings can offer technological support to reduce severe crop yield loss after delay or disease control interventions.