Skip to content
NTPU AI Learn

AI 301 · Module 8

Multi-Agent Architecture Canvas

From an agent team to an architecture you can build

Start Section 1
  • Duration45 min
  • LevelArchitect
  • PrerequisiteAI 202 recommended, plus Modules 1 to 7 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 Multi-Agent Architecture Canvas and what each one records
  • Follow the recommended Canvas flow from Shared Goal to Final Output
  • Connect the Canvas sections to the deliverables from Modules 2 to 7
  • Answer the final reflection question about the value of each additional Agent and the risk of coordination overhead
  • List the seven deliverables of AI 301 and connect them to the Canvas

Goal, reasons, roles and boundaries

9 min

The Canvas, part 1: Shared Goal, Why Multi-Agent?, Agent Roles, Responsibility Boundaries

The Multi-Agent Architecture Canvas is the main artifact of AI 301. It puts the goal, the agents, their responsibilities, how they communicate, how they are controlled and how they are evaluated 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 should multiple Agents divide work, communicate, delegate, coordinate and complete a shared goal?

Canvas sections 1 to 4
#SectionWhat it records
1Shared GoalWhat the multi-agent system should accomplish
2Why Multi-Agent?Why one agent is insufficient: specialization, permission separation, parallel work, independent verification, organizational boundaries or scalability
3Agent RolesFor each agent: Role, Purpose, Responsibilities, Capabilities, Knowledge, Tools, Permissions, Inputs and Outputs
4Responsibility BoundariesFor each role: what it Owns, what it May Assist, what it Must Not Do and what Requires Approval

Info

Human, AI, Human: a person defines the goal and the boundaries, the agents work as a team inside them, and a person reviews and remains responsible for the result.

Key takeaway: Start with a clear shared goal, a real reason for more than one agent, and roles with boundaries.

Allocation, delegation, communication and context

9 min

The Canvas, part 2: Task Allocation, Delegation Rules, Communication Protocol, Shared Context & Memory

Sections 5 to 8 describe how work moves and how information is shared. They bring in Modules 3 to 5.

Canvas sections 5 to 8
#SectionWhat it records
5Task AllocationChoose: static, rule-based, supervisor-selected, capability-based or dynamic
6Delegation RulesSender, receiver, task, context, constraints, expected output and completion criteria
7Communication ProtocolMessage types, required fields, evidence, status, error handling and escalation
8Shared Context & MemoryShared context, private context, shared knowledge, restricted information, the source of truth and who owns updates

Info

Your Delegation & Handoff Map from Module 3 fills sections 5 and 6. Your Agent Communication Contract from Module 4 fills section 7. Your Shared Context Map from Module 5 fills section 8.

Key takeaway: Sections 5 to 8 say who does which task, how it is passed, what messages carry and who sees what.

Orchestration, conflict, oversight and evaluation

9 min

The Canvas, part 3: Orchestration Pattern, Handoff Points, Conflict Resolution, Human Oversight, Evaluation

Sections 9 to 13 describe how the whole team is controlled and how you know it works. They bring in Modules 6 and 7.

Canvas sections 9 to 13
#SectionWhat it records
9Orchestration PatternChoose and justify: Supervisor / Worker, Planner / Executor, Specialist Team, Peer-to-Peer, Hierarchical or Hybrid
10Handoff PointsWhere responsibility moves from one agent to another
11Conflict ResolutionRule-based resolution, evidence comparison, evaluator or supervisor, and human escalation
12Human OversightWhere a person reviews, approves, overrides or receives an escalation
13EvaluationSuccess metrics: task quality, coordination, latency, cost, duplication, conflict rate and escalation rate

Info

Your Orchestration Pattern Design from Module 6 fills sections 9 and 10. Your Multi-Agent Evaluation & Escalation Policy from Module 7 fills sections 11 to 13.
Check yourself

Which Canvas sections does your Multi-Agent Evaluation & Escalation Policy from Module 7 mainly fill?

Which Canvas sections does your Multi-Agent Evaluation & Escalation Policy from Module 7 mainly fill?

Key takeaway: Sections 9 to 13 hold your orchestration design and your Evaluation & Escalation Policy.

Risks, output and the flow

9 min

The Canvas, part 4: Risks & Boundaries, Final System Output, and the flow

Canvas sections 14 and 15
#SectionWhat it records
14Risks & BoundariesRole confusion, circular delegation, context leakage, duplicated work, conflicting goals, excessive communication and uncontrolled cost
15Final System OutputWhat the agent team ultimately produces
Recommended Canvas flow
  1. Shared Goal
  2. Why Multi-Agent?
  3. Define Roles
  4. Assign Responsibilities
  5. Allocate Tasks
  6. Delegate
  7. Communicate
  8. Share / Isolate Context
  9. Coordinate
  10. Conflict? If no, Continue; if yes, Resolve or Escalate
  11. Orchestrate
  12. Evaluate
  13. Final Output

Info

Text version of the diagram: start from the Shared Goal and ask Why Multi-Agent? Then define roles, assign responsibilities, allocate tasks, delegate, communicate and decide what context is shared or isolated. Next, coordinate. At that point ask whether there is a conflict: if not, continue; if so, resolve it or escalate. Then orchestrate, evaluate, and finish with the Final Output.
Check yourself

In the recommended flow, what happens after the agents coordinate and a conflict is found?

In the recommended flow, what happens after the agents coordinate and a conflict is found?

Key takeaway: List the risks explicitly, and follow the flow from the Shared Goal to the Final Output, handling conflict on the way.

Complete your Canvas

9 min

Complete your Canvas, and where AI 301 leads

Use your own task from the course, or a simulated one, to complete the Canvas. Reuse your work from Modules 1 to 7 and fill the gaps.

Final reflection question

What clear value does each additional Agent provide, and how will the system prevent coordination overhead from becoming greater than that value?

  • Agent Role Cards
  • Delegation & Handoff Map
  • Agent Communication Contract
  • Shared Context Map
  • Orchestration Pattern Design
  • Multi-Agent Evaluation & Escalation Policy
  • Multi-Agent Architecture 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 and the boundaries, the agent team works inside them, and a person reviews and remains responsible.
  • M5276 Agentic AI Innovative Applications, Week 13: this course supports the case discussion on multi-agent collaboration, role division, communication protocols, agent-to-agent delegation and orchestration.
  • AI 202, Agentic AI Planning & Reasoning: asks how one agent decides what to do next. AI 301 asks which agent should do the work and how several agents coordinate, and it does not repeat AI 202.
  • AI 302, Agent Harness Engineering & Reliable Automation: a future course that goes deeper into reliable execution and operation. AI 301 does not teach it.

Key takeaway: Your Canvas is a reviewable design, and its key test is whether each extra agent earns its coordination cost.

Save as PDF

What next?