Subjects › AGI Agents (6-week course)
W4 · AutoGen & CrewAI in practice
What the frameworks package, sequential vs hierarchical crews, guardrails, and counting the cost.
The idea
AutoGen AgentChat packages group chat (round-robin or selector) with termination conditions; the course notes it is now in maintenance mode, with Microsoft Agent Framework as the successor. CrewAI packages roles, tasks and processes: sequential runs tasks in order, passing earlier outputs as context; hierarchical adds a manager that delegates — at the cost of extra LLM calls. Flows wrap deterministic Python steps around a crew when you need control and reproducibility (BriefFlow). Exercise 4.5's guardrail is a deterministic Python check on a task's output: it returns (True, output) to accept or (False, feedback) to make the agent retry. Compare systems on accuracy and cost, never accuracy alone. In your repo: Week04_MultiAgent_AutoGen_CrewAI/autogen_team/, crewai_crew/ (pipelines.py, guardrails.py, integration.md); Exercise 4.5.
Lesson
No lesson built yet. The worked solutions below are the material it would be assembled from.
Questions (5)
| Question | Type | Difficulty | Drift | Accuracy |
|---|---|---|---|---|
| You want your AutoGen team to speak in a fixed, predictable order — planner, then researcher, then executor — … | mcq | Easy -1.20 | — | — (0) |
| On the milestone's 5 tasks, System A (a single ReAct agent) gets 4 correct using 3 LLM calls per task. System … | numeric | Moderate -0.30 | — | — (0) |
| Exercise 4.5 adds numbers_are_sourced, a deterministic guardrail on a CrewAI task's output. Why implement a ch… | mcq | Moderate 0.30 | — | — (0) |
| The crew in pipelines.py has three tasks: research, analysis and brief. Use this simplified cost model (an ass… | numeric | Hard 1.00 | — | — (0) |
| BriefFlow in pipelines.py runs kb_search as a plain Python step first, then hands the facts to a sequential cr… | short | Very hard 1.50 | — | — (0) |