Learn AI /AI Agents and Automation /Planning
Planning
- Intermediate
- AI Agents and Automation
- Technical readers
- Policy officers
Also called: task decomposition.
Back to the milestone you came from
In five seconds
What it is
Planning is the process of turning a goal into ordered sub-tasks, checks and possible actions so a system can act step by step.
In plain words
Planning means turning a goal into smaller steps, doing them in order, and changing course if new information appears.
An everyday example
Amina receives a group message: “Samira told Lina that they need to move the meeting because the office is closed.” Before replying, Amina does not jump to one answer. She checks earlier messages, notes who organised the meeting, sees who mentioned the office, and lists two possible meanings of “they”. Then she asks one short clarifying question and only after that updates the calendar note.
This analogy shows careful step-by-step work, but it can mislead in two ways. First, a person may use broad world knowledge and social context that an AI system may not have unless those facts are provided. Second, AI planning in software often includes explicit tool calls, machine-readable state, and automatic retries, which do not appear in a simple chat example.
Remember thisPlanning turns a broad goal into ordered actions and decision points. A capable planner also updates those actions when new information appears, while respecting permissions, budgets and approval boundaries.
This term is contested
The broad idea is widely used, but boundaries between planning, reasoning, orchestration and workflow vary across research papers and products.
Understand it
Understand it
Why it matters
Planning affects what an AI system actually does, not just what it says. If a system can break work into steps, call tools and revise its next action, it can be useful for scheduling, research support and routine administration. It can also create new risks: wasted tool calls, hidden failure points, unsafe actions and overconfidence if checks are weak. For managers and policy staff, this matters when choosing approval gates, audit logs, access controls and when a human must review or stop the process.
How planning turns a goal into actions
After viewing, the reader should be able to explain that planning breaks a goal into sub-tasks, checks results after each step, and may revise the next action.
01
1. Goal and limits
The system receives a goal plus constraints such as approved tools, deadlines or required human review.
Draw a box labelled Goal and limits with small tags underneath: tools, rules, time, approvals.
02
2. Break into sub-tasks
The system drafts smaller steps, such as collect data, compare options, ask a question, then act.
Draw a second box labelled Sub-tasks with a short stacked list icon.
03
3. Take a step
The system performs the next step, often by calling a tool or producing a message.
Draw a third box labelled Action or tool call with a small wrench or connector icon.
04
4. Check observation
The result of the step is checked. Missing data, errors or new facts may change what should happen next.
Draw a diamond-shaped checkpoint labelled Observation check connected from the action box.
05
5. Finish or revise
If the goal is met, the process ends. If not, the system updates the remaining steps and continues, or asks for human input.
Draw two outgoing arrows from the checkpoint: one to a Finish box, one looping back to Sub-tasks with label Revise.
A left-to-right step diagram with five boxes connected by arrows. It starts with a goal and constraints, moves to sub-task selection, then tool or action steps, then a checkpoint based on observations, and ends with either completion or revision of the plan. Read from left to right to see how a plan is made and updated.
On hover or tap, each step highlights and reveals one sentence of explanation; selecting the checkpoint highlights the loop-back path to show replanning.
Describe this without the diagram
The diagram presents planning as five ordered stages. First, a system receives a goal and limits such as allowed tools, rules, time or approval requirements. Second, it breaks the goal into smaller sub-tasks. Third, it performs one step, often using a tool. Fourth, it checks the observation returned from that step. If the observation shows success, it finishes. If the observation shows missing information, an error or a changed situation, it revises the remaining sub-tasks or asks for human input before continuing.
With reduced motion: Show the same five stages as a static numbered sequence with a visible loop arrow from Observation check back to Break into sub-tasks, and plain text labels for Finish and Revise.
In AI agents and automation, planning means breaking a goal into smaller actions before and during execution. The input is usually a goal, some instructions, available tools, and limits such as time, cost or approval rules. A tool is an external system the agent can call, such as a search service, calendar or database.
What happens next is not magic. The system proposes a sequence of sub-tasks, such as gather information, compare options, ask for missing details, then perform an allowed action. It may revise that sequence after each observation, meaning new information returned from a tool or a previous step.
The output is usually a plan or a choice of the next action. Sometimes the plan is written out in text. Sometimes it stays internal and only the selected actions and final result are visible. Good planning also includes checks, such as whether a required field is missing or whether human approval is needed.
Planning is often confused with reasoning models, workflows and tool use. A reasoning model uses extra computation during answering to improve difficult tasks. Tool use means calling external software. A workflow is an ordered set of steps that moves work from input to result. Planning specifically chooses or adjusts steps for a goal and situation, and may operate within a workflow.
How “Finding who “they” refers to in a messy message thread” maps onto the real thing
| In the story | In the AI system |
|---|---|
| Amina’s goal of replying correctly | The agent’s overall task |
| Checking earlier messages for clues | Gathering observations before acting |
| Listing two possible meanings | Breaking the task into candidate sub-steps or hypotheses |
| Asking a clarifying question before changing the calendar | Revising the plan and adding a safe checkpoint before action |
A worked example
A reporting officer in a country office asks an approved assistant to assemble the clearance package for the weekly situation report by close of business.
Starting from: Goal: prepare the draft and supporting file. Available tools: the document repository, an internal task tracker and a drafting tool. Constraints: use only cleared inputs, preserve source links and do not circulate anything without approval.
- The assistant plans the work: retrieve the current template, locate cleared section inputs, identify gaps, check figures and dates against their sources, assemble the draft, then prepare it for clearance.
- After checking the repository, it finds that one programme section is missing. It updates the plan rather than drafting around the gap and prepares a reminder for the responsible focal point.
- Once the missing input appears in the tracker, it retrieves the cleared version, adds it to the draft and runs consistency checks across the summary, tables and source links.
- It flags two unresolved discrepancies for the reporting officer and stops before circulation, leaving the draft, source list and clearance checklist together in the file.
Result: The reporting officer receives a nearly complete clearance package with its sources preserved, two issues requiring human judgement clearly flagged, and no unauthorised circulation. The plan changed when new information became available but respected the approval boundary.
In a United Nations workplace
During the technical evaluation of three bids for a regional protection information platform, a protection officer in a regional bureau uses an approved internal assistant to organise the workflow. The assistant plans the sequence from conflict-of-interest confirmations and compliance checks through criterion-by-criterion review, clarification questions and completion of the technical evaluation report by close of business.
The planner turns the deadline into ordered tasks, tool calls and decision points. It can track which checks are complete, revise the plan when a required document is missing, and prepare the file for the panel. It must stop at technical judgement, scoring and sign-off, which remain with the evaluators.
Before you try thisBid documents, evaluators' comments and personal data in submitted CVs are procurement-sensitive and must not be entered into any tool not approved for that material. Survivor, witness or case data must never be used to test bidders' claims. Access should be limited to the evaluation panel, and the assistant must not score bids, contact bidders or submit the report autonomously.
Where you may hear it
- A vendor demo during a procurement meeting
- An ICT governance board review of an agent pilot
- A design session for an assistant connected to internal tools
- An evaluation office discussion on testing agent behaviour
Avoid confusion
Avoid confusion
Often confused with
Workflow
A workflow is an ordered set of steps that moves information or tasks from input to result, often combining software, people and checks.
Reasoning model
A reasoning model is a language model designed for stronger performance on multi-step tasks, often using extra computation while answering, though visible steps may not reflect its internal process.
Tool use
A system uses external software or devices during a task to look up information, calculate, carry out actions or control equipment beyond its own processing.
Common misconceptions
Common misconception: Planning means the system always creates a full plan before doing anything.
Correction: Often it does not. Many systems plan a little, act, inspect the result, then update the next step. This matters because real tasks change when new information arrives, tools fail, or a user clarifies the goal.
Common misconception: If a system can plan, it can be trusted to act on its own.
Correction: Planning only organises steps. It does not guarantee the goal is appropriate, the facts are correct, the tool output is reliable, or the action is safe. High-impact tasks still need permissions, monitoring and human review.
Common misconception: Planning is the same as a fixed workflow.
Correction: A workflow is usually predesigned: step A then B then C. Planning is more flexible. It chooses, orders or revises sub-tasks for the current situation, though in practice many systems combine both approaches.
Common misconception: Good text output proves good planning.
Correction: A system may describe a sensible plan in words but still call the wrong tool, skip a check or fail to update after new observations. Planning quality should be assessed from behaviour and outcomes, not fluent wording alone.
Quick check
Which example best shows planning rather than only tool use?
The second option shows a goal being decomposed into sub-tasks, with an information check and an approval step. That is planning. The first is only tool use, the third is memory, and the fourth is a fixed workflow rather than context-specific planning.
Go deeper
Go deeper
The simple version
Planning means turning a goal into smaller steps, doing them in order, and changing course if new information appears.
The technical version
In agentic systems, planning is the mechanism that converts a high-level objective into an executable sequence or policy over sub-tasks, tool calls and decision points. Inputs can include the user goal, system instructions, available APIs or connectors, prior observations, memory, and operational constraints such as budgets, permissions and approval rules. The planner may produce an explicit plan, an implicit next-action choice, or a rolling plan that is updated after each observation. In practice, planning often interacts with tool use, short-term memory, orchestration and feedback loops. It is distinct from tool use itself, because selecting and ordering actions is different from merely having access to a tool. It is also distinct from a workflow itself, because a workflow moves work through ordered steps, while planning selects or revises actions for a goal. Some systems rely on a language model to generate sub-goals or evaluate next steps; others use classical search, rule-based methods, or hybrid approaches. The quality of planning is usually judged by task completion, efficiency, robustness to failures, and whether the system respects constraints and escalation rules.
Reviewed external resources
Primary sourceModel Context Protocol (opens in a new tab)Model Context Protocol
This official Model Context Protocol material shows how tools and context can be connected in a standard way, which suits technical readers building or evaluating agent systems.
Deeperlink checked 2026-07-31
EncyclopaediaAutomated planning and scheduling (opens in a new tab)Wikipedia
This Wikipedia article gives a broad overview of automated planning and scheduling, which suits readers who want historical context and standard terminology.
Intermediatelink checked 2026-07-31
ArticleBuilding effective agents (opens in a new tab)Anthropic
This Anthropic article explains agents in practical terms and helps readers who want to see how planning fits into tool-using systems.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Planning (opens in a new tab)YouTube search
This YouTube search results page helps visual learners find different video explanations of planning, but it is a search page rather than one vetted lesson.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Planning (opens in a new tab)Reddit search
This Reddit search results page shows community discussions and examples of how people talk about planning in practice, but it is a search page and suits readers who want informal perspectives.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
Planning turns a broad goal into ordered actions and decision points. A capable planner also updates those actions when new information appears, while respecting permissions, budgets and approval boundaries.