Skip to content
NTPU AI Learn

AI 202 · Module 7

Agent Planning & Reasoning Canvas

Turn planning and reasoning into a system you can design

Start Section 1
  • Duration45 min
  • LevelAdvanced
  • PrerequisiteAI 104 recommended, plus Modules 1 to 6 of this course
0 / 5 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

  • Name the 15 sections of the Agent Planning & Reasoning Canvas and what each one records
  • Follow the recommended Canvas flow from Goal to Stop, Escalate or Complete
  • Link the Tools section to AI 201 without duplicating the MCP Tool Canvas
  • Answer the key reflection question about the next action and the evidence that would change the plan
  • List the six deliverables of AI 202 and connect them to the Canvas

Goal, state and subgoals

9 min

The Canvas, part 1: Goal, Initial State, Observations, Subgoals

The Agent Planning & Reasoning Canvas is the main artifact of AI 202. It puts an agent goal, its state, its actions, its reasoning, its evaluation and its re-planning on one page, so the design can be reviewed and handed to the next stage. It has 15 sections. This slide covers the first four.

Core question

How do we design an agent goal, state, actions, reasoning, evaluation and re-planning into an architecture that can be built?

Canvas sections 1 to 4
#SectionWhat it records
1Agent GoalThe final objective, what counts as success, and what is explicitly out of scope
2Initial StateWhat the agent knows at the start: context, memory, known constraints and available resources
3ObservationsWhat the agent can observe while running: user feedback, tool results, retrieved knowledge, environment state and intermediate outputs
4SubgoalsHow the main goal is decomposed into smaller goals

Info

Human, AI, Human: a person defines the goal and its limits, the agent plans and acts, and a person reviews and remains responsible for the result.

Key takeaway: Start the Canvas with a clear goal, a known starting state, the observations you can rely on, and the subgoals.

Actions, strategy and decisions

9 min

The Canvas, part 2: Actions, Tools, Planning Strategy, Decision Policy

Sections 5 to 8 describe what the agent can do and how it chooses. They connect directly to Modules 2 to 4.

Canvas sections 5 to 8
#SectionWhat it records
5Available ActionsWhat the agent can do, for example reason, retrieve, ask, search, call a tool, generate, compare, verify, escalate or stop
6ToolsWhich actions depend on AI 201 tools. For each tool: its purpose, when it may be used, relevant permissions and the result expected
7Planning StrategyChoose or combine: Fixed Plan, Rolling Plan, ReAct, Search, Tree Search, the MCTS concept, Adaptive Planning
8Decision PolicyFor each major decision: what is being decided, what state or evidence is needed, what options exist and which criteria choose the next action

Warning

The Tools section refers to AI 201. It lists which actions rely on a tool and under what conditions, and it does not duplicate the MCP Tool Canvas or describe tool schemas.
Check yourself

What belongs in the Tools section of the Planning & Reasoning Canvas?

What belongs in the Tools section of the Planning & Reasoning Canvas?

Key takeaway: Sections 5 to 8 say what the agent can do and how it chooses. Tools point to AI 201 and do not repeat it.

Evaluation, reflection and boundaries

9 min

The Canvas, part 3: Evaluation, Reflection, Re-planning Triggers, Constraints

Sections 9 to 12 describe how the agent knows whether it is on track and what limits it. They bring in Modules 5 and 6.

Canvas sections 9 to 12
#SectionWhat it records
9EvaluationHow the agent knows an action worked. Dimensions: correctness, completeness, relevance, cost, latency, risk, source quality, user acceptance
10ReflectionWhen the agent should reflect. What failed? Was the action wrong, the plan wrong or the observation incomplete? Is more evidence needed?
11Re-planning TriggersWhich conditions cause the plan to change
12Constraints & BoundariesAction limits, data restrictions, cost limits, step limits, time limits and prohibited actions

Info

Your Evaluation & Reflection Policy from Module 5 fills sections 9 and 10. Your Re-planning Policy from Module 6 fills sections 11 and 12.

Key takeaway: Sections 9 to 12 hold your Evaluation & Reflection Policy and your Re-planning Policy.

Escalation, stopping and output

9 min

The Canvas, part 4: Escalation, Stop Conditions, Final Output, and the flow

Canvas sections 13 to 15
#SectionWhat it records
13Human EscalationWhen the agent must involve a person, for example low confidence, a high-impact action, an irreversible action, an ambiguous objective, conflicting instructions or a policy exception
14Stop ConditionsSuccess stop, failure stop, safety stop, budget stop and human takeover
15Final OutputThe artifact or outcome the agent should produce
Recommended Canvas flow
  1. Goal
  2. Initial State
  3. Observe
  4. Reason
  5. Plan
  6. Choose Action
  7. Tool / Execute
  8. Observe Result
  9. Evaluate
  10. Reflect? If no, Continue
  11. If yes: Re-plan? If no, Continue; if yes, make a New Plan
  12. Stop, Escalate or Complete

Info

Text version of the diagram: the agent starts from the Goal and Initial State, then loops: Observe, Reason, Plan, Choose Action, Tool or Execute, Observe Result, Evaluate. After Evaluate it asks Reflect? A no continues the loop. A yes leads to Re-plan? A no continues, and a yes creates a New Plan. The loop ends when the agent decides to Stop, Escalate or Complete.
Check yourself

In the recommended flow, what happens right after the agent evaluates a result and reflection says the plan is wrong?

In the recommended flow, what happens right after the agent evaluates a result and reflection says the plan is wrong?

Key takeaway: The flow is a loop with reflection and re-planning inside it, and it ends only through Stop, Escalate or Complete.

Complete your Canvas

9 min

Complete your Canvas, and where AI 202 leads

Use your own task from the course, or a simulated one, to complete the Canvas. Reuse your work from Modules 1 to 6 and fill the gaps.

Key reflection question

Why should this agent choose this next action, and what evidence would cause it to change its plan?

  • Goal & Plan Map
  • ReAct Trace
  • Search / Candidate Action Map
  • Evaluation & Reflection Policy
  • Re-planning Policy
  • Agent Planning & Reasoning Canvas (the primary artifact)

Warning

Privacy reminder: use a simulated or already-public task. Do not enter company secrets, client personal data or unauthorized sensitive information in the Canvas.

Done

Human, AI, Human: you define the goal, the agent plans and acts inside the boundaries, and a person reviews and remains responsible.
  • M5276 Agentic AI Innovative Applications, Week 10: this course supports the case discussion on planning and reasoning engines.
  • AI 301, Multi-Agent Collaboration & Orchestration: a future course that goes deeper into role division and delegation. AI 202 does not teach it.
  • AI 302, Agent Harness Engineering & Reliable Automation: a future course that goes deeper into reliability and recovery. AI 202 does not teach it.

Key takeaway: Your Canvas is a reviewable design, and its key test is why this next action, and what evidence would change the plan.

Save as PDF

What next?