The paper "TwoSTEP: Multi-agent Task Planning using Classical Planners and Large Language Models" addresses the challenge of multi-agent planning by combining classical planning techniques with large language models (LLMs). Classical planning, such as PDDL, guarantees a sequence of actions to achieve a goal state, but it does not capture temporal aspects of action execution. In contrast, LLMs can leverage commonsense reasoning to assemble action sequences, but they do not guarantee execution success. The authors propose TwoSTEP, a method that decomposes a multi-agent planning problem into two single-agent planning problems, where one agent acts as a helper and the other as the main agent. The helper agent's subgoal is derived using LLMs, which are guided by human intuition and domain knowledge. This decomposition allows for faster planning times and more efficient execution steps compared to solving the multi-agent problem directly. The paper evaluates TwoSTEP on both symbolic and simulated domains, demonstrating that it achieves shorter execution lengths and comparable planning times to single-agent planning while maintaining execution success. Additionally, the LLM-inferred subgoals are found to approximate those specified by human experts. The results show that TwoSTEP effectively leverages LLMs to decompose multi-agent tasks, leading to more efficient and successful planning.The paper "TwoSTEP: Multi-agent Task Planning using Classical Planners and Large Language Models" addresses the challenge of multi-agent planning by combining classical planning techniques with large language models (LLMs). Classical planning, such as PDDL, guarantees a sequence of actions to achieve a goal state, but it does not capture temporal aspects of action execution. In contrast, LLMs can leverage commonsense reasoning to assemble action sequences, but they do not guarantee execution success. The authors propose TwoSTEP, a method that decomposes a multi-agent planning problem into two single-agent planning problems, where one agent acts as a helper and the other as the main agent. The helper agent's subgoal is derived using LLMs, which are guided by human intuition and domain knowledge. This decomposition allows for faster planning times and more efficient execution steps compared to solving the multi-agent problem directly. The paper evaluates TwoSTEP on both symbolic and simulated domains, demonstrating that it achieves shorter execution lengths and comparable planning times to single-agent planning while maintaining execution success. Additionally, the LLM-inferred subgoals are found to approximate those specified by human experts. The results show that TwoSTEP effectively leverages LLMs to decompose multi-agent tasks, leading to more efficient and successful planning.