Multi-Agent Systems quiz

24 questions from the Multi-Agent Systems track. Score 80% or higher to unlock a certificate.

Why Multi-Agent?
1. When is adding a second agent justified?
- Always, because teams are more intelligent
- When tools, context, or success checks actually diverge
- When you need more tokens on the marketing site
- When the first prompt is already perfect
Split or Merge
2. Two personas can both search and both stop when the paragraph “feels done.” What should you do?
- Add a third persona named Director
- Merge them — they are one agent
- Give both refund tools
- Run them in a swarm
The Coordination Tax
3. What is the coordination tax?
- Extra GPUs
- Schema, who-speaks, memory access, team evals, and a loop cap — paid for every extra agent
- A longer system prompt
- More thumbs-up in chat
The Interface Is Data
4. What should a researcher hand a coder?
- The full web dump plus thoughts
- A typed brief with citation ids
- A poem about the bug
- Direct access to run_shell
Keep a Single-Agent Baseline
5. The team matches the single agent on quality and costs 5×. What should you do?
- Add more personas
- Keep the baseline (or prove a different win: safety, latency, incidents)
- Hide the cost in another budget
- Debate every FAQ
Roles: Planner, Worker, Critic
6. Why should the critic usually not have write tools?
- Critics are not smart enough to write
- Otherwise you have two workers, and the critique path disappears into extra side effects
- Frameworks forbid it
- Write tools only work in single-agent mode
Tool Isolation Is the Point
7. The planner wants to patch src/app.py to save a hop. What should happen?
- Allow it — planners are senior
- Deny — path is not on the planner allow-list
- Give the planner every tool
- Ask the critic to patch instead
The Critic Does Not Write
8. Tests are red. What should the critic do?
- Call edit to save time
- Return a fix list and reassign the worker
- Refund the customer
- Delete the tests
A Blackboard, Not a Group Chat
9. Where should the current plan live?
- In the last chat message
- On a typed blackboard (store) keyed by the run
- Only in the planner’s thoughts
- In every worker’s system prompt as a novel
Handoffs Are Typed Events
10. What is a legal planner-to-worker handoff?
- A chat paragraph
- A typed event with to, step id, inputs, and a budget
- The planner’s full transcript
- Direct run_shell from the planner
Orchestration: Sequential, Handoff, Supervisor
11. Which orchestration pattern makes workers unable to talk to each other?
- Handoff with unlimited hops
- Sequential with a shared scratchpad novel
- Supervisor (dispatcher) assigns work and collects results
- A group chat with no schema
Sequential Is the Default
12. Who should choose the next node in a sequential pipeline?
- The worker who feels inspired
- Your code (the pipeline list)
- A 70B model on every hop
- The customer
Hop Limits and Acyclic Graphs
13. Two agents keep handing the same ticket back and forth. What is the first control to add?
- A third agent named Manager with more tools
- A hop limit (and an acyclic graph or a supervisor that owns finish)
- Longer personas
- Removing all stop conditions
Supervisor Star, Not a Mesh
14. A coder sends a private note to the docs worker to skip tests. What should the runtime do?
- Deliver it — they are on the same team
- Drop it — peer messages are forbidden in supervisor mode
- Promote it to the system prompt
- Refund the customer
Supervisor Plus a Sequential Subgraph
15. Who may skip HITL on a billing subgraph?
- The billing worker, for VIPs
- Nobody — HITL is in the pipeline list, not a thought
- The critic
- A swarm child
Debate and a Judge
16. What makes a debate judge useful?
- Always picking the longer answer
- A rubric grounded in evidence, tests, or citations — plus a round cap
- Using the same prompt as the proposer
- Letting the opponent write to production
Ground the Judge
17. Two models agree, with no citations. What should the judge do?
- Ship the chorus
- Fail closed or handoff — agreement is not evidence
- Pick the longer one
- Launch a swarm
Swarms of Cheap Workers
18. What is the main risk of giving every swarm worker write tools?
- They will type faster
- Races, duplicate side effects, and an unbounded blast radius
- Reduce becomes too accurate
- Cheap models refuse to write
Reduce Is the Product
19. How should you merge 200 swarm children into the parent prompt?
- Paste every trace
- Reduce to a table (vote, deduped ids, dropped/missing) — then the parent sees the table
- Pick the longest child
- Ask a 70B model to “summarize the vibe”
A Write Barrier After Map
20. When may a swarm child call refund?
- When its thought is confident
- Never — refund waits for the parent apply after reduce
- If two children agree
- After three map retries
How Multi-Agent Systems Fail
21. A swarm of 50 large children would cost more than the job cap. What should launch do?
- Launch anyway and hope
- Refuse to start — price before launch
- Hide the cost in another budget
- Let each child refund the customer
Detect Ping-Pong
22. A and B bounce the same brief three times. What should the runtime return?
- One more hop
- cannot: ping-pong (or cycle) — then handoff
- A longer persona
- finish with a guessed answer
One Writer Per Record
23. If two roles can mutate the same customer record, what do you have?
- A high-performing team
- A distributed race with a chat UI
- A swarm
- A grounded judge
When Not to Swarm
24. You have 200 independent tickets to score, a reduce table, and a priced cap. What should you ship?
- A debate on every ticket
- A swarm (map-reduce) with a write barrier
- Fifty editors on the same file
- One group chat with 50 personas