An Agent Framework for Land Resource Monitoring and Analysis

This paper proposes an agent framework for land resource monitoring and analysis. The framework organizes and encapsulates domain knowledge and operational workflows into standardized skill units, enabling the agent to perform task planning, execution, and reflection under skill-based constraints and thereby improving the procedural consistency of complex operational tasks. In addition, a task-semantics-based result verification module is introduced to assess the semantic consistency of intermediate results and further identify latent errors in which tools execute successfully but the resulting outputs deviate from user requirements. Experimental results show that the proposed method outperforms the ReAct, Plan-and-Execute, and Plan-and-React baselines overall, achieving a final result accuracy of 93.33% ± 1.44%. Ablation experiments demonstrate that skill-based constraints, progressive loading, and result verification all improve agent execution reliability. A dedicated evaluation of the result verification module further shows that it can effectively identify latent execution errors, achieving a Precision of 93.55%, a Recall of 90.63%, and a false-positive rate of 6.25%.