What is Agent Planning?
Also called planning, AI planning.
Agent planning is the process by which an agent works out an ordered set of steps to reach a goal before, or while, carrying them out. Plans may be produced once at the start, revised after each observation, or replaced entirely when new information arrives. Planning quality is a common bottleneck, because a confident plan built on a wrong assumption fails at every step.
Two broad styles exist. Plan-and-execute writes the full sequence up front and then runs it, which is efficient and easy to inspect but fragile when reality differs from the assumption. Interleaved planning decides only the next step, using what the last result showed, which adapts well but can wander. Many systems keep a coarse plan and re-decide the details each step.
Plans fail for predictable reasons. The agent assumes a tool returns data it does not have. It orders steps so a later one needs output an earlier one never produced. It plans past the point where information exists, inventing steps for a situation it cannot yet see. Explicit replanning after each result, and short plans rather than long ones, mitigate most of this.
Planning is hard to evaluate because a plan can be reasonable and still fail, or unreasonable and still succeed. Useful signals are whether each step names a real tool, whether preconditions are satisfied in order, and how often the agent revises. Judging only the final answer hides a system that reached the right result by an expensive and unreliable route.
Planning depends on decomposition to produce the units and on state to know what has been done. In classical AI, planning meant search over formally defined actions with guaranteed properties. Language model planning offers no such guarantees, which is why plans are checked by execution and by validators rather than trusted on the strength of how sensible they read.
Key points
- Plans can be written up front or decided one step at a time.
- Most plan failures come from unmet preconditions and wrong assumptions.
- Frequent replanning beats long plans made in advance.
- A readable plan is not evidence of a workable one.
In practice
Asked to move a project's open items into a new tracker, an agent first plans: list items, create the destination, copy each one, verify counts. On listing, it finds attachments it did not expect, so it revises the plan to upload files before creating each item. Without that revision every copy step would have silently dropped the attachments.