Accelerating climate change and intensifying human activities are reshaping the resilience of ecological networks, especially in arid and semi-arid agro-pastoral transition zones where ecosystems remain highly disturbance-sensitive. Long-term resilience dynamics in northwestern China's agro-pastoral ecotone (1985-2023) were analyzed by constructing ecological networks using an integrated "importance-sensitivity connectivity" framework combined with self-organizing map neural networks and circuit theory. Five disturbance scenarios were applied to assess temporal resilience trajectories and identify effective optimization strategies. Three key findings emerged. First, resilience exhibited a distinct U-shaped evolution across three phases: initial stability (1985-2000), mid-term fluctuation (2000-2015), and restabilization (2015-2023), closely mirroring major policy transitions. Second, resilience responses were disturbance-specific, indicating differentiated adaptive capacities. Networks remained robust under random attacks but vulnerable to targeted disturbances, while resistance to overgrazing improved most substantially under long-term grazing management. Third, optimization strategies exhibited scenario-dependent thresholds with paradoxical effects: source expansion enhanced sensitivity-based resilience by 14.68% but reduced overgrazing resilience by 10.25%; corridor optimization peaked at 0.1-0.2 addition ratios before triggering negative returns. These results support a spatial-temporal-functional framework to guide disturbance-specific strategies across management zones, clarifying mechanisms shaping resilience evolution in arid ecosystems and providing a transferable basis for precision ecological engineering and resilience-oriented land management globally.