LLM Agent Planning

LLM agent planning studies how language models can turn instructions into sequences of actions that achieve goals in tools, software, websites, and physical or simulated environments. Common approaches include ReAct-style plan-and-act loops, hierarchical subgoals and skill graphs, symbolic planners such as PDDL, search methods such as tree search, and world models that predict action outcomes; reinforcement learning and trajectory feedback are also used to improve these systems. Evaluations span web and GUI navigation, coding, household tasks, games, and tool-use benchmarks, increasingly testing long horizons, changing conditions, tool failures, and resource costs. A central challenge is that a plausible plan does not guarantee faithful or safe execution: agents must track state, ground tool arguments, verify results, and recover when assumptions or plans fail.