Agent planning

Agent planning is the stage in which an AI agent breaks a given goal into a sequence of ordered steps and decides which tools it will use, and in what order. Rather than reacting blindly to a single prompt, the agent first sketches a path to the goal and then carries it out step by step, revising the plan as the results of each action change the picture.

Planning differs from a rigidly defined agentic workflow: in a workflow, a human designs the order of steps in advance, whereas in planning the agent itself sets the sequence based on the goal and the current context. Planning is tightly linked to tool use — a plan is only an intention, and the agent achieves real results by calling tools: search engines, APIs, or code.

In deployment, the quality of planning decides whether an agent can handle a multi-step task or gets stuck after the first unexpected result. That is why planning usually comes with controls: an explicit record of the plan, the ability to review it, and points where a human approves the next steps.