Autonomous Agent Architecture for Complex Tasks via Hierarchical Planning and Language Model Reasoning

Abstract

This study presents a planning-capable autonomous agent framework based on large language models. The framework addresses key problems in existing agents, including broken reasoning chains, unstable planning structures, and inconsistent strategy execution in complex tasks. It treats the large language model as the core reasoning engine and constructs task semantic representations, hierarchical subgoal structures, and multi-stage reasoning mechanisms. This enables the agent to perform task understanding, goal decomposition, and strategy generation within a unified structure. A multi-stage semantic state update mechanism is introduced to maintain stability during transitions between subtask stages. A hierarchical task decomposition module maps global instructions into ordered and executable subgoal sequences, which improves the controllability and precision of strategy generation. Experiments on multiple goal-oriented tasks evaluate success rate, path efficiency, subgoal accuracy, and planning completion. The results show clear improvements in execution quality and multi-stage reasoning, confirming the importance of combining structured planning with language-based reasoning. Sensitivity analyses on planning depth, action generation temperature, and environmental visibility further reveal the behavioral patterns of the agent under different parameter settings. Overall, the framework integrates semantic understanding, hierarchical planning, and behavior generation and provides a unified and complete design path for building stable and highly structured autonomous agents.