planning

advancedspreading

Planning is the ability of an AI agent to break down a larger goal into steps, order them, and adjust as needed during execution.

In plain terms

Think of a weekend grocery shopping trip and cooking. You do not rush to the store impulsively: first you decide what to cook, check what you have at home, write a list, and figure out the best order to proceed. An AI agent plans similarly: it creates a sequence of steps from the given goal. Some write out the entire plan in advance, while others only figure out the next one or two steps and proceed along the way.

This distinguishes the agent from a simple chatbot. A chatbot gives one answer to one question. An agent receives a task requiring ten or a hundred interdependent steps, which it must figure out itself, for example, searching first, then comparing, and only then booking.

A good plan is not rigid. If there is no sour cream in the store, you buy something else instead of going home empty-handed. A well-built agent also replans when a step fails or finds something different than expected.

How it works

Definition of goal: the agent clarifies what the final result should be and what conditions must be met (deadline, budget, prohibited operations).
Breakdown: it splits the large task into smaller, individually executable subtasks and determines which builds on which.
Tool selection: assigns to each step what will be used to perform it, such as web search, opening a file, calculation, or involving another agent.
Execution and observation: it performs the steps in sequence and checks the result after each one.
Replanning: if something does not go as expected, it adjusts the remaining steps and, if necessary, stops and asks you for clarification.

An everyday example

You ask the agent to organize a three-day autumn trip to Eger for four people within a given budget. Before booking anything, it shows you its plan: checking available dates, searching for accommodation, compiling activities, summing up costs, and only booking after your approval. Seeing this list, you can speak up in time if you want something different.

Who is affected

employeeIf you entrust a multi-step task to an agent, you can assess in advance whether it is heading in the right direction based on its plan. → Work & HR

entrepreneur, executiveThe quality of planning determines whether an agent can be entrusted with an entire workflow or only individual subtasks.

developerThe coding agent decides which files to access and in what order based on the plan, so it is advisable to review the plan before writing code. → Code

everyoneThe plan is the point where one can still intervene cheaply and without risk in what the machine is about to do.

A common misconception

MythIf the agent writes a convincing, detailed plan, it will execute it exactly as written.

In realityA plan is only an intention: the agent may deviate from it mid-process, skip a step, or build on a faulty assumption. The final result must always be checked separately.

Limits and risks

A plan can be plausible yet flawed because the agent builds it on an unverified assumption.
In long tasks, errors accumulate: one early wrong step can lead all subsequent steps astray.
Design takes time and money; for simple tasks it may be an unnecessary slowdown.
Approving the plan is not enough: before irreversible steps (payment, deletion, sending an email), it is advisable to have the step itself approved by a human, because the agent may deviate from the plan.

Similar, but not the same

reasoning – Reasoning involves thinking through a question to reach an answer, while planning determines the sequence of future actions.
orchestration – Orchestration refers to the division of labor and coordination between multiple agents or tools, while planning determines what needs to be done and in what order.
munkafolyamat (workflow) – In a workflow, humans write and record the steps in advance, whereas in planning, the agent devises them itself for the given task.

Related concepts

Broader

AI-agent, agentic AI, artificial intelligence

Narrower

task decomposition, redesign, plan mode

On the edge

reasoning, chain of thought, tool use, orchestration, human approval

A brief history

Machine planning is one of the oldest research areas in artificial intelligence: the STRIPS planning system, which calculated a robot's steps, was described at the Stanford Research Institute in 1971. Planning for agents built on large language models was widely popularized by the ReAct study published in 2022, which demonstrated the alternation of reasoning and action.

Where things stand

At the end of 2026, planning is a basic capability for most agents and often appears in coding agents as a separate, approvable planning mode. For long, multi-step tasks, its reliability is still variable, so human oversight remains necessary.

Status: 2026 November

In the news

Appears in 44 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Data centers accounted for at least 60% of US electricity demand growth last year, grid struggles to keep upNov 5, 2026 · Ipar
Gemini in Google Classroom offers teachers planning, quiz, rubric and text-levelling toolsNov 2, 2026 · Education & Family
AI teacher training: Rwanda passes 90,000, Mongolia sets strategy, Kenya plans trainingOct 27, 2026 · Education & Family
US data-centre gas turbines could erase 15 years of emissions cuts, report saysOct 27, 2026 · Ipar
2027 B2B marketing budgets need a line for GEO/AEO, planning guides sayOct 26, 2026 · Marketing
Data center boom spurs grid modernization; New England utility uses AI tools to set prioritiesOct 26, 2026 · Ipar
Andhra Pradesh makes AI training compulsory for 80,000 state school teachers by month-endOct 19, 2026 · Education & Family
Eurobarometer: more and more European companies use artificial intelligenceOct 16, 2026 · Ipar

Also known as

planning · AI planning · agent planning · task planning · Task planning · automated planning · plan mode · agent design

Sources

Yao et al.: ReAct: Synergizing Reasoning and Acting in Language Models (arXiv, 2022)

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.