From Plan Generation to Reliable Execution: An Architecture-Centric Survey of Embodied Long-Horizon LLM Agents

Long-horizon embodied agents employ large language models (LLMs) or vision-language models (VLMs) as cognitive cores to complete multi-stage tasks through sustained interaction. Their primary challenge is not to generate long plans, but to maintain task state under partial observability, environmental change, and delayed feedback, adapt actions to execution outcomes, and prevent local errors from propagating. Research is accordingly moving from plan generation and skill invocation toward state maintenance, execution verification, and failure recovery. Existing surveys, however, commonly organize the literature by foundation model or individual technique and thus provide limited insight into how these mechanisms jointly shape system behavior. We propose an architecture-level framework spanning planning, action interfaces, memory, execution feedback, world models, and multi-agent collaboration. A Method-by-Dimension Matrix compares module information flow, invocation conditions, and authority to modify behavior, and we assess whether current evaluations support architectural attribution. The evidence indicates that long-horizon capability depends less on plan length, memory capacity, or module count than on maintaining task- and environment-consistent state and ensuring that module outputs affect subsequent behavior when required. We synthesize the resulting trade-offs, open problems, and research directions to provide common coordinates for method comparison, architecture design, and evaluation.