Planning and decomposition in multi-step agent tasks

Two main approaches: (1) one-shot plan: have the LLM produce a full DAG of subtasks upfront, then execute (good for known-shaped tasks, brittle if assumptions break); (2) iterative re-plan: execute one subtask, observe result, re-plan based on what you learned (more robust but slower). Common techniques: least-to-most prompting (decompose from abstract to concrete subquestions), tree-of-thoughts (branch and evaluate multiple plans), self-consistency (sample N plans, take majority), "plan-and-execute" loops where a planner LLM refines a plan as observations come in. Senior nuance: in 2026 production, isolated step-by-step monolith agents are giving way to (a) graph-of-thought DAG executors (LangGraph) and (b) constraint-checking plans: plans that include assertions ("if A fails, retry with strategy B").