Researchers propose Pheromone-Guided Policy Optimization (PhGPO) to improve long-horizon tool planning for Large Language Model (LLM) agents, leveraging historical trajectories to guide policy optimization and improve tool transitions.