LLM-based agents demonstrate impressive capabilities in language understanding and reasoning but often falter when facing complex tasks involving domain-specific rules. This limitation primarily stems from the lack of explicit modeling of expert knowledge and dependency constraints, leading to illogical subtask sequences, planning hallucinations, and structural deficiencies in domain knowledge. To address this, we propose the knowledge-enhanced task decomposition (KTD) framework, which integrates knowledge graphs with chain-of-thought reasoning. By constructing dual graphs that encode both domain constraints and expert knowledge, the KTD provides structured guidance for complex task decomposition, enabling the generation of semantically complete, logically coherent, and executable subtask sequences. Experiments on three multi-hop QA datasets (2WikiMultiHopQA, HotpotQA, and MuSiQue) demonstrate that the KTD consistently outperforms mainstream prompting methods across various Qwen2.5 model sizes. Notably, in the low-resource 1.5B setting, the KTD achieves up to a 13.86% accuracy improvement, confirming its effectiveness in mitigating planning hallucinations and addressing structural knowledge deficiencies.