TWOSTEP is a method for multi-agent task planning that combines classical planning with large language models (LLMs). It decomposes a multi-agent planning problem into two single-agent planning problems, leveraging LLMs to infer subgoals for individual agents that can be executed in parallel. The method uses a helper agent to generate subgoals in English, which are then translated into PDDL format for classical planning. The helper's subgoals are designed to reduce the plan execution steps for the main agent, allowing for faster and more efficient execution. The main agent then uses these subgoals to complete the remaining tasks. TWOSTEP ensures execution success by relying on classical planning guarantees, while also achieving faster planning times and shorter execution lengths compared to traditional multi-agent PDDL planning. The method was evaluated on symbolic domains and a simulated environment, showing that it achieves comparable or better performance in terms of planning time and execution length. Results indicate that TWOSTEP can effectively decompose tasks into parallel subgoals, leading to more efficient multi-agent planning. The method also demonstrates that LLM-inferred subgoals can approximate those specified by human experts, highlighting the potential of combining LLMs with classical planning for complex task planning scenarios.TWOSTEP is a method for multi-agent task planning that combines classical planning with large language models (LLMs). It decomposes a multi-agent planning problem into two single-agent planning problems, leveraging LLMs to infer subgoals for individual agents that can be executed in parallel. The method uses a helper agent to generate subgoals in English, which are then translated into PDDL format for classical planning. The helper's subgoals are designed to reduce the plan execution steps for the main agent, allowing for faster and more efficient execution. The main agent then uses these subgoals to complete the remaining tasks. TWOSTEP ensures execution success by relying on classical planning guarantees, while also achieving faster planning times and shorter execution lengths compared to traditional multi-agent PDDL planning. The method was evaluated on symbolic domains and a simulated environment, showing that it achieves comparable or better performance in terms of planning time and execution length. Results indicate that TWOSTEP can effectively decompose tasks into parallel subgoals, leading to more efficient multi-agent planning. The method also demonstrates that LLM-inferred subgoals can approximate those specified by human experts, highlighting the potential of combining LLMs with classical planning for complex task planning scenarios.