Definition
Agent planning turns a goal into intermediate steps, tool calls, or checkpoints. A plan may be produced before execution or revised as new observations arrive. Some systems lay out several future steps, while others plan only the next intermediate step.
The application or agent runtime controls what information reaches the model and decides which requested actions to execute. A generated plan is data used by the agent loop, not an authority that can bypass tool permissions or workflow rules.
Simple example
A repository agent receives the goal “update the service to the new configuration format”. It first proposes to locate the configuration model, find its consumers, modify the parser, update examples, and run targeted tests. After discovering a second configuration source, it revises the remaining steps before making further changes.
Each file read, edit, and test invocation still happens through application-provided tools with their own validation and limits.
Why it matters
Planning can make multi-step work easier to inspect and can reduce aimless tool use. Explicit intermediate goals give the application useful places to apply budgets, request approval, record progress, or stop when assumptions become invalid.
Planning is especially useful when later actions depend on information gathered earlier. It is less useful when the workflow is already known and can be expressed reliably as deterministic code.
One important nuance
A plausible plan is not evidence that the steps are correct, complete, or safe. Models can invent dependencies, plan from stale assumptions, or continue following a plan after observations contradict it. Treat plans as revisable proposals. Validate tool calls at execution time, re-evaluate after important results, and keep authorization, termination, and side-effect controls outside the model.