Abstract: Objectives Current ship navigation decision-making systems struggle to demonstrate superior performance in undefined sailing scenarios. Given the broad applicability of large language models (LLMs) in unknown scenarios, this study proposes a dual-LLM-core-driven adaptive ship navigation agent architecture (Nav-DLLC). Methodology Nav-DLLC firstly uses a large-parameter LLM as the control core with ReAct-based prompting to decompose navigation tasks and invoke external tools, reducing LLM errors. Subsequently, a small-parameter LLM fine-tuned with low-rank adaptation serves as the collision avoidance core, processing unstructured data to generate COLREG-compliant decisions. Results Simulation experiments demonstrate that Nav-DLLC achieves outstanding performance in both traditional ship collision avoidance tasks and unstructured dynamic scenarios. Its collision avoidance accuracy reaches 86%, and its behavior compliance rate reaches 90%, significantly outperforming LLM baselines and traditional methods such as the Dynamic Window Approach (DWA) and Artificial Potential Field (APF). The decision core’s single-decision latency is 11.13 seconds, higher than the 0.73 seconds of traditional methods, yet remains within the safe time window for collision avoidance decision-making. Conclusion Nav-DLLC bridges the gap between traditional navigation systems and LLM technology, providing a safe and efficient intelligent decision-making paradigm for complex navigation environments.