PrerequisiteAI 202 (Agentic AI Planning & Reasoning) recommended; AI 201 Tool and Permission concepts helpful
0 / 5 completed
Sync my progressJoin a session to keep progress across devices and share it with your instructor
Learning objectives
Distinguish a Single Agent, Agent + Tool, Agent + Human and a Multi-Agent System
Apply the decision model: One Goal, can one agent complete it safely and effectively, and if not, why not
Name the five reasons that can justify a multi-agent design: expertise, permissions, parallel work, independent verification and organizational separation
Weigh the costs of a multi-agent design: complexity, coordination cost, latency and reliability
Explain separation of duties, capability boundaries and context boundaries
Choose one real use case, described with simulated data, to carry through the whole course
Problem and need
8 min
When one agent is asked to do everything
In AI 202 you worked with one agent that plans, reasons and adapts. This course asks the next question: what if the work is too large, too risky or too varied for one agent? The core question is: when is one agent not enough, and when does a multi-agent design only add complexity?
Work scenario (simulated)
One agent is asked to prepare a research grant application: check it against the funder rules, build the budget, write the narrative, review its own work and mark it ready to submit. It does all five well enough, but nobody can tell whether the review was independent, because the same agent that wrote the text also approved it. It also had access to the budget system and the submission form at the same time, which nobody had intended.
Very different kinds of expertise in one prompt tend to make each part shallower
One agent holding every permission means one mistake can reach everything
Independent steps run one after another, although they could run in parallel
The agent that produces the work also judges it, so there is no independent check
Warning
The opposite mistake is just as common: splitting a simple task across several agents because it sounds more advanced. Every extra agent adds messages, waiting time and new ways to fail. Start from the goal, not from the architecture.
Key takeaway: A multi-agent design is a response to a specific problem, not an upgrade. First ask whether one agent can do the job safely and effectively.
Core concepts
10 min
Four architectures: from Single Agent to Multi-Agent
Between one agent and a team of agents there are simpler options. Always try the simplest architecture that meets the goal, and move to the next one only when you can name what the simpler one cannot do.
From Single Agent to Multi-Agent System
Single Agent: one agent handles the whole task with its own knowledge and context
Agent + Tool: one agent, with tools for search, calculation or data access (AI 201)
Agent + Human: one agent, with a person who decides, approves or supplies what the agent cannot
Multi-Agent System: several agents with different roles, working toward one shared goal
Text equivalent: the four architectures run in a line from the simplest to the most complex. A Single Agent needs no coordination. An Agent + Tool adds capabilities but still has one decision maker. An Agent + Human adds a person for judgment and accountability. A Multi-Agent System adds several agents, so it also adds the work of coordinating them.
Comparing the four architectures (simulated examples)
Architecture
What it adds
Example
Single Agent
Nothing beyond the agent itself
Summarize one meeting transcript into action items for a person to review
Agent + Tool
Access to data or calculation
Answer staff questions by searching the approved leave policy documents
Agent + Human
Human judgment, approval or missing information
Draft a reply to a student grade query; a teacher approves before it is sent
Multi-Agent System
Separate roles that can specialize, run in parallel or check each other
Separate agents draft, check compliance and review a grant application, under one shared goal
Check yourself
An agent answers staff questions by searching the approved leave policy documents. Which architecture is the simplest that fits?
Key takeaway: Use the simplest architecture that meets the goal. Move up only when you can say what the simpler one cannot do.
12 min
The decision model: why not one agent?
A multi-agent design needs a reason. The decision model turns "should we use several agents?" into a short sequence of questions, and it defaults to one agent.
Decision model: One Agent or Many?
One Goal: state a single shared goal
Can one agent complete it safely and effectively? If yes, prefer a Single Agent (with a tool or a person if needed)
If not, ask: why not?
Name the reason: different expertise, different permissions, parallel work, independent verification or organizational separation
Only then consider a Multi-Agent design
Text equivalent: the model starts from one goal and asks whether a single agent can complete it safely and effectively. If the answer is yes, stop and prefer the Single Agent. If the answer is no, you must name why not. There are five accepted reasons. Only when at least one of them applies do you consider a multi-agent design.
Five reasons that can justify a multi-agent design (simulated examples)
Reason
The question to ask
Example
Different expertise
Do the parts need genuinely different knowledge or instructions?
Legal compliance rules and budget arithmetic each need their own focused instructions
Different permissions
Should some steps be allowed to do things others must not?
The agent that gathers quotes must not be able to approve a purchase
Parallel work
Are there independent parts that can run at the same time?
Check five documents at once instead of one after another
Independent verification
Must a check be done by someone other than the producer?
A reviewing agent that did not write the draft checks it against the rules
Organizational separation
Do different teams own different data, rules or responsibilities?
Finance and the research office each keep their own data and approval rules
Info
Reasons are not a score to add up. One strong reason, such as a real permission boundary, can be enough. Five weak reasons are not. If you cannot state the reason in one sentence, prefer one agent.
Check yourself
A team wants three agents to write, translate and proofread a one-page notice. Nothing needs different permissions, checks are done by a person, and the pieces are small. What does the decision model suggest?
Key takeaway: Default to one agent. A multi-agent design needs a named reason: expertise, permissions, parallel work, independent verification or organizational separation.
Costs and boundaries
10 min
The costs, and what separation buys
Every extra agent has a price. Before you add one, compare what you gain with what you will pay, using the same words each time.
The price of a multi-agent design
Cost
What it means
Example (simulated)
Complexity cost
More parts to design, explain and review
Four roles need four sets of instructions and four sets of tests
Coordination cost
Agents must exchange information, and they can misunderstand each other
The reviewing agent needs the draft, the rules and the context, and any of them can be missing
Latency
Waiting for handoffs and replies
Three sequential agents can take three times as long as one
Reliability
More links means more places to fail
If one agent returns a wrong result, the next one may build on it
What separation can buy
Specialization: each agent has a focused role and instructions
Capability boundaries: each agent has only the tools it needs
Context boundaries: each agent sees only the information it needs
Separation of duties: the producer is not also the approver
What it costs
More messages and more things that can be misread
Longer time to the first useful result
A harder system to test, explain and review
Roles that overlap or leave gaps if they are poorly designed
Info
Human-AI-Human: a person defines the goal and decides which parts are worth separating, the agents carry out their roles, and a person reviews the result and stays responsible for it. Splitting work across agents does not split away accountability. Later modules cover delegation, communication and orchestration. Durable execution, checkpoints and recovery are beyond this course and belong to AI 302.
Check yourself
Which statement best describes separation of duties in a multi-agent design?
Key takeaway: Separation buys specialization, boundaries and independent checks, and costs complexity, coordination, latency and reliability. A person stays responsible.
Practice and apply
10 min
Practice: One Agent or Many?
Activity 1 gives you 8 simulated use cases. For each, choose one: Single Agent, Agent + Tool, Agent + Human or Multi-Agent, and justify the architecture with the decision model. Some cases can reasonably be read two ways; what matters is that you can explain your reasoning. Please do not enter real company secrets, client personal data or any unauthorized sensitive information in this exercise.
8 use cases (decide for yourself)
Case
Use case
Think about
1
Summarize one meeting transcript into action items for a person to review
Is there any reason to separate?
2
Answer staff questions from the approved leave policy documents
What does the agent need to reach the documents?
3
Draft a reply to a student grade query; a teacher must approve before it is sent
Who holds the final decision?
4
Review a grant application: check funder rules, check the budget and check the writing, in parallel; the checker must not be the drafter
Which of the five reasons apply?
5
Turn a table of course enrollments into a chart
Is this a capability or a role?
6
Prepare a purchase recommendation: the agent gathering quotes must not be able to approve payment, and the payment data sits in another department
Permissions? Organizational separation?
7
Shorten a paragraph of a report
What would a second agent add?
8
Sort incoming support messages and pass urgent ones to the duty officer
What is the person for? What is the tool for?
Single Agent: one agent can complete it safely and effectively on its own
Agent + Tool: one agent is enough, but it needs a tool for data or calculation
Agent + Human: one agent is enough, but a person must judge, approve or supply information
Multi-Agent: at least one of the five reasons applies and the benefit outweighs the coordination cost
1Apply: choose one real use case for the whole course Pick a task that looks complex enough that one agent might not be sufficient: several kinds of expertise, different permissions or a separate check. Every later module practices on it, and it ends up in your Multi-Agent Architecture Canvas.
2Describe it with simulated data Write only the goal and the nature of the work, with no real names, client personal data or company secrets. If the task involves sensitive content, use a simulated version.
3Write the goal and the first decision One sentence for the shared goal, then run it through the decision model: can one agent do it safely and effectively? If not, which of the five reasons applies?
You need to pick one use case for the whole course. Which is the best fit?
Key takeaway: Human-AI-Human: a person defines the goal and the architecture, the agents carry out their roles, and a person reviews and stays responsible for the result.
