Planning

Two main approaches

• Plan-and-execute: have the model produce a full step list first, then execute item by item. Clear and reviewable, but inflexible when reality diverges from the plan.
• Interleaved (e.g. ReAct): do not list everything up front; take a step, then look. Adaptive, but prone to "forgetting" the overall goal midway.

Many systems compromise: produce a rough plan as a skeleton, then allow local rewrites during execution based on observations.

Why it decides long-task success

Long tasks usually fail not because one step is done wrong but because the decomposition itself is flawed — a missing precondition, wrong ordering, too-coarse granularity. Good planning surfaces risk early, while it is still cheap. Methods like Tree of Thoughts go further, expanding several candidate plans in parallel and selecting.

Common misconceptions

• A plan is not one-shot: good agents re-plan when evidence changes rather than forcing the original.
• A more detailed plan is not always better; over-decomposition amplifies cumulative per-step error.

Related terms

Sources

1. Wang et al. — Plan-and-Solve Prompting (arXiv:2305.04091)
2. Yao et al. — Tree of Thoughts (arXiv:2305.10601)
3. Russell & Norvig, AIMA (4th ed.), part on Planning