agent planning

Planning is the agent skill of looking at a big, vague goal and breaking it into an ordered list of smaller steps it can actually carry out. 'Plan a weekend trip to Tainan' becomes: check the calendar, find trains, book a hotel near the old town, make a short food list. Without planning, an agent tends to lunge at the goal in one leap and flail; with it, the work becomes a sequence it can tackle one piece at a time.

There are two broad styles. In up-front planning, the model writes the whole step list first and then executes it; in interleaved planning, it plans only the next step, acts, sees the result, and re-plans. Up-front plans are easy to inspect but go stale when reality differs; interleaved planning adapts but can wander. Many systems combine them: a rough high-level plan, plus step-by-step adjustment underneath.

Plans from an LLM are suggestions, not guarantees. The model can produce a clean-looking plan with a missing step, an impossible ordering, or a task no available tool can actually do — which is why planning is rarely done once and then trusted. In practice plans are checked, revised, and often regenerated as the run unfolds and reality pushes back.