Agent Planning — AI Glossary

Reference details

Related terms

• AI Agent
• Reasoning Model
• Orchestrator Agent

How it works

1. The goal is decomposed into sub-tasks with their dependencies, which is the first step and not the whole activity.
2. The sub-tasks are ordered against what the agent can actually do, since a plan containing a step with no corresponding tool cannot execute.
3. A planning strategy is chosen: the whole plan up front, one step at a time, or a hierarchy where high-level steps expand later.
4. Execution reports back, and the plan is revised when a step returns something it did not anticipate — this is the part that distinguishes planning from decomposition.
5. Termination is a judgement against the goal, so it needs an external cap on steps and spend to stop a plan that keeps revising itself.

Example: The trade between planning up front and planning as you go

One-shot planning makes a single planning call, so a ten-step task costs 1 planning call plus 10 executions: 11. Step-by-step planning re-plans after every result: 10 planning calls plus 10 executions, about 21 — nearly double. The second is roughly twice the cost and is the only one that can react when step three reveals the plan was wrong. Hierarchical planning sits between: plan three phases up front, expand each when reached, so about 4 planning calls. The choice is how likely the plan is to survive contact.

Commonly confused with

Planning is the whole activity — decomposing a goal, ordering the pieces, and revising when a step returns something unexpected. Task decomposition is the first of those, and the decomposition entry says as much itself. The distinction is worth keeping because most agent failures attributed to bad planning are actually the missing third part: a decomposition was produced once and never revisited when reality diverged from it.

When to use it

Reach for it when:

• Goals need several steps whose order is not fixed by code in advance.
• Steps may reveal information that should change the remaining work.
• Progress must be reportable against named sub-goals mid-run.

Reach for something else when:

• The sequence is always the same, where a fixed workflow is cheaper and testable.
• Single-step tasks, where planning adds a round trip and nothing else.
• No cap on steps or spend, leaving a self-revising plan to set its own budget.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Agent Planning — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path