Skip to content
NTPU AI Learn

AI 202 · Module 2

Goal Decomposition & Planning Fundamentals

How does an agent turn a goal into a plan it can actually carry out?

Start Section 1
  • Duration55 min
  • LevelAdvanced
  • PrerequisiteAI 202 Module 1 recommended; AI 104 and AI 201 recommended
0 / 6 completed
Sync my progressJoin a session to keep progress across devices and share it with your instructor

Info

After you join, your completed lesson numbers are sent to your instructor together with your learner code, so they can see your progress. Use the codes your instructor gave you, and do not enter your name, company secrets or other personal data. Progress is kept with the rest of that session's data.

Learning objectives

  • Decompose a Goal into Subgoals, a Candidate Plan, Actions and Expected Outcomes
  • Explain the difference between designing a workflow (AI 104) and an agent choosing or revising a plan (AI 202)
  • Use planning questions, including which actions are reversible and what would invalidate the plan
  • Identify Preconditions, Constraints and Dependencies of an action
  • Choose a suitable planning horizon and judge whether a plan is still valid
  • Produce a Goal & Plan Map for your own use case using simulated data

Problem and need

8 min

A goal is not yet a plan

In Module 1 you saw that an adaptive agent chooses its next step from the situation. But choosing a next step well needs something to aim at. "Explain why sign-ups dropped" is a goal, not a plan: it does not say what to do first, what to do next, or how to know progress is being made. The core question of this module is: how does an agent turn a goal into a plan it can carry out?

Work scenario (simulated)

A planning assistant is told: "Prepare the options for a new room booking system." It starts reading product pages at once. Two hours later it has 40 open notes, no clear comparison criteria, and it never checked what the department is actually allowed to buy. The work was not wrong, but it had no plan: no subgoals, no order, no idea of what "done" looks like.

  • Without subgoals, effort goes to whatever is easiest to start, not what matters most
  • Without dependencies, work is done in the wrong order and redone
  • Without success criteria, nobody can tell whether progress is real
  • Without knowing which actions are reversible, a small mistake can become an expensive one

Warning

The opposite mistake is planning too much: a very detailed plan for a situation that keeps changing is quickly wrong. A good plan is detailed enough to act on and light enough to revise.

Key takeaway: A goal says where to go. A plan says what to do first, what next, and how to tell that it is working.

Core concepts

9 min

From Goal to Subgoals to Actions

Planning breaks a goal into smaller pieces until each piece is something the agent can actually do and check. Five layers describe the path from the goal down to what should happen.

The planning model
  1. Goal: the result to achieve
  2. Subgoals: smaller results that together achieve the goal
  3. Candidate Plan: a proposed order of work to reach the subgoals
  4. Actions: the concrete steps the agent can take
  5. Expected Outcomes: what should be true after each action, so progress can be checked

Text equivalent: the five layers run from top to bottom. The Goal is split into Subgoals. A Candidate Plan proposes how to reach them, using Actions. Each Action has an Expected Outcome, which is what the agent later compares against the actual Observation. It is called a candidate plan because it is a proposal that may be revised.

Example: room booking options (simulated). Goal: a short comparison a manager can review
LayerExample
GoalA one-page comparison of three room booking options that a manager can review
SubgoalsKnow the requirements; know the budget and rules; find candidate options; compare them
Candidate PlanConfirm requirements first, then check rules, then search options, then compare
ActionsRead the requirement notes; check the purchasing guideline; search product pages; fill a comparison table
Expected OutcomesA list of five requirements; a budget limit; three options; a table with one row per option
Check yourself

Which item is an Expected Outcome rather than an Action?

Which item is an Expected Outcome rather than an Action?

Key takeaway: Every action in a plan should have an Expected Outcome, so the agent can tell whether it worked.

10 min

Designing a workflow vs. choosing a plan

You already know how to decompose a task from AI 104. The difference is who produces and revises the plan, and when.

AI 104: design the workflow

  • A person designs the steps, order and checkpoints in advance
  • The path is fixed before the run
  • Good when the path can be written down
  • The question is how to design the flow well

AI 202: the agent chooses or revises the plan

  • The agent proposes a plan from the goal and the current state
  • The plan may change during the run as observations arrive
  • Good when the path cannot be fully predefined
  • The question is how the agent chooses or revises the plan, within limits

A useful set of planning questions keeps this disciplined. The agent, or the person setting up the agent, should be able to answer all of them before acting.

Eight planning questions
QuestionWhy it matters
What is the goal?Everything is judged against it
What state are we in?The plan depends on what is already known and done
What constraints apply?Limits on data, time, cost, permissions and privacy
What actions are available?A plan can only use actions that exist and are allowed
What must happen first?Dependencies decide the order
Which actions are reversible?Reversible actions can be tried; irreversible ones need more care and often a person
What evidence indicates progress?Defines the Expected Outcomes to check
What would invalidate the plan?Names the changes that would force a revision

Info

Human-AI-Human: a person sets the goal and constraints, the agent proposes and revises the plan inside them, and a person approves plans that include irreversible or high-impact actions.
Check yourself

What is the clearest difference between AI 104 workflow design and AI 202 planning?

What is the clearest difference between AI 104 workflow design and AI 202 planning?

Key takeaway: AI 104 designs the path in advance. AI 202 lets the agent choose or revise it, inside limits set by a person.

9 min

Preconditions, Constraints, Dependencies and reversibility

A plan is only as good as the actions in it. For each action, check four things before the agent takes it.

Four checks for every action
CheckMeaningExample (simulated)
PreconditionsWhat must already be true before the action can workTo compare options, the requirement list must exist
ConstraintsLimits the action must respectUse only the approved product list; no personal data; within the budget
DependenciesWhich other actions or results this one waits onThe comparison table needs the search results and the budget limit
ReversibilityCan the action be undone or safely retried?Reading a page is reversible; sending the comparison to a supplier is not
  • Put reversible, low-cost actions early: they teach the agent about the situation cheaply
  • Put irreversible or external actions late, and only after a person has approved
  • If a precondition is not met, the action should not be chosen yet
  • A missing dependency is a reason to reorder the plan, not to guess

Warning

Do not treat a confident-sounding plan as a checked plan. The agent can propose a plan that ignores a constraint or an unmet precondition. A person, or a stated check, should confirm the plan before actions with real impact.
Check yourself

A plan has two actions: "read the department purchasing guideline" and "email three suppliers for quotes". How should they be ordered and handled?

A plan has two actions: "read the department purchasing guideline" and "email three suppliers for quotes". How should they be ordered and handled?

Key takeaway: Check Preconditions, Constraints, Dependencies and reversibility for every action, and put irreversible ones last with a person approving.

Horizon and validity

9 min

Planning horizon and plan validity

How far ahead should the agent plan? That is the Planning horizon. Plan too far, and the plan is full of guesses. Plan too near, and the agent wanders without direction.

Short-horizon planning

  • Plan the next one to three actions only
  • Cheap to revise after each observation
  • Suits uncertain or fast-changing situations
  • Risk: locally good steps that do not add up to the goal

Long-horizon planning

  • Plan many steps ahead toward the goal
  • Gives direction and shows dependencies early
  • Suits stable situations with well-known steps
  • Risk: a long plan built on assumptions that turn out wrong

In practice, agents often combine both: keep the goal and subgoals fixed as the long view, plan the next few actions in detail, and revise after each observation.

Signs that a plan may no longer be valid (simulated examples)
SignExampleWhat to do
An assumption turned out falseThe budget limit in the notes is out of dateStop and confirm with a person before continuing
A precondition can no longer be metThe required data source is unavailableChoose another action or ask for access
An observation contradicts the planTwo sources disagree on a key requirementRevise the plan or escalate; do not silently pick one
The goal or constraint changedThe manager narrows the scope to two optionsUpdate the plan against the new goal
Check yourself

The situation changes often and each observation may change what matters. Which approach fits best?

The situation changes often and each observation may change what matters. Which approach fits best?

Key takeaway: Plan far enough to keep direction and near enough to revise cheaply, and know what would make the plan invalid.

Practice and apply

10 min

Practice: Goal, Subgoal, Action

Activity 2 asks you to take a realistic task and turn it into a plan. Use the example below to practice, then do the same for your own use case from Module 1. Please do not enter real company secrets, client personal data or any unauthorized sensitive information in this exercise.

Worked example (simulated): prepare a briefing on options for a new room booking system
ItemExample
GoalA one-page comparison of three options that a manager can review this week
SubgoalsRequirements known; budget and rules known; options found; options compared
Candidate ActionsRead requirement notes; read purchasing guideline; search product pages; build comparison table
ConstraintsApproved products only; no personal data; do not contact suppliers without approval
Success CriteriaThree options, each checked against all five requirements and the budget, with sources listed
Reversible?All reading and drafting steps are reversible; contacting suppliers is not, and needs approval
What would invalidate the plan?A change in the budget limit or requirements; the guideline is unavailable
  1. Write your Goal
    One sentence for your use case: the result, who reviews it, and by when.
  2. List Subgoals and Candidate Actions
    Break the goal into three to five subgoals and list one or two actions for each. Order them by dependency.
  3. Add Constraints and Success Criteria
    Write the limits (sources, privacy, time) and how you will know each subgoal is done.
  4. Mark reversibility and invalidation
    Mark which actions are irreversible or external and need approval, and write what would invalidate your plan.

Info

Deliverable: the Goal & Plan Map, a single page with Goal, Subgoals, Candidate Actions, Constraints and Success Criteria. It becomes part of your Agent Planning & Reasoning Canvas.
Check yourself

Which statement best describes a good Goal & Plan Map?

Which statement best describes a good Goal & Plan Map?

Key takeaway: Human-AI-Human: a person sets goal and limits, the agent proposes and revises the plan, and a person approves irreversible actions and reviews the result.

Save as PDF

What next?