Tool-Use Planning

Tool-use planning studies how AI agents, especially large language models, select tools and organize their calls into workflows that accomplish a goal. Approaches include representing tool dependencies as graphs, searching over possible plans with methods such as Monte Carlo tree search, and training on successful trajectories or execution feedback. Benchmarks test agents on API workflows, typed tool interfaces, and grounded settings such as geographic or physical tasks, where correct tool choice alone is not enough: plans must also bind parameters and integrate results. A central challenge is balancing reliable, generalizable long-horizon plans against execution cost, especially as tool catalogs and task constraints change.