Task planning is important in intelligent proxy systems to help agents turn advanced goals into effective operations. But now methods often face problems with problem expansion, modeling dependency, and open environment stability. To address these problems, this article offers a new planning and planning structure that combines large language models (LM) with graphical reasoning. After providing a description of tasks in natural language, LLM will disassemble it into simple subtains, as well as determine local priority relationships, which will then become global dependencies. The results of the subtaspecty were compared, causing targeted self-correcting. Multi-round verifiability tests ensured the reliability of the problem of the sub-tasting task. Experiments on a multi-examination dataset set of 100 samples showed that this method worked much better than the baseline, and reflection-based approach, indicating that it has the potential to create stable and interpreted intelligent agents.
