Abstract
To address the issues of insufficient goal parsing, loose stage connections, insufficient consistency in the execution chain, and weak constraint satisfaction in the autonomous planning and execution process of large language model agents in complex task scenarios, this paper proposes an autonomous planning and execution method based on task graph construction and stage dependency constraint modeling of large language model agents. This method first jointly encodes user goals, contextual environment, and historical interaction information to form a unified semantic representation for complex tasks, and then generates and structures multi-stage sub-tasks based on this representation. Subsequently, by constructing a task graph representation containing stage nodes, dependencies, and constraint semantics, the hierarchical structure and logical connections within the complex task are explicitly depicted. Furthermore, a stage dependency constraint modeling mechanism is introduced to uniformly constrain preconditions, state transmission, and execution feasibility, thereby enhancing the overall coherence of the planning process and the stability of the execution process. During the execution phase, a stage-aware strategy is used to dynamically select and control the sequence of currently executable nodes, enabling the model to maintain strong structural consistency and decision rationality in complex task processing. The comparative results show that the proposed method achieves superior performance in terms of task completion capability, planning integrity, execution consistency, and constraint satisfaction. This demonstrates that the proposed framework can effectively improve the autonomous planning and execution quality of large language model agents in complex task scenarios, providing new methodological support for the construction of highly reliable intelligent agents for real-world application environments.
