Subjects › AGI Agents (6-week course)
W4 · Multi-agent coordination
Roles, blackboards, turn-taking, critics — and when adding agents helps or just adds cost.
The idea
A blackboard is shared memory every agent reads and writes; roles (planner, researcher, executor, critic) specialise prompts and tools. Turn-taking can be round-robin or model-selected. In collaboration_loop.py the critic's APPROVE ends the loop and REVISE starts another full round, up to max_rounds; in Exercise 4.4 the AutoGen critic's 'APPROVE. TERMINATE' becomes the only way to end the chat, with MaxMessageTermination as a backstop. A critic that uses the same model as the executor shares its blind spots. 'Adding agents instead of fixing tools or prompts' heads the course's list of anti-patterns (TUTOR_NOTES.md); the Week 4 README says to start every design by asking whether one agent with better tools would do. In your repo: Week04_MultiAgent_AutoGen_CrewAI/collaboration_loop.py, autogen_team/; Exercise 4.4 (make_critic).
Lesson
No lesson built yet. The worked solutions below are the material it would be assembled from.
Questions (5)
| Question | Type | Difficulty | Drift | Accuracy |
|---|---|---|---|---|
| In collaboration_loop.py the planner, researcher, executor and critic never call each other. How do they coord… | mcq | Easy -1.30 | — | — (0) |
| Your team's researcher keeps calling kb_search with badly formed queries and misses facts. A teammate proposes… | mcq | Easy -0.60 | — | — (0) |
| In collaboration_loop, each round gives four roles one turn in order: planner, researcher, executor, critic. T… | numeric | Moderate 0.30 | — | — (0) |
| The AutoGen team in team.py stops on TextMentionTermination("TERMINATE") | MaxMessageTermination(12), and the … | mcq | Hard 1.00 | — | — (0) |
| In collaboration_loop the critic and the executor use the same model. Can the critic reliably catch the execut… | short | Very hard 1.50 | — | — (0) |