Exercise 2: Multi-Agent Role Allocation (CrewAI)¶

Pairs with Stage 4 — Agent Frameworks Exercise 2.

Task¶

Three agents each own one step, collaborating to produce a blog intro:

Researcher → Writer → Critic
  (find facts)  (write)  (verify, PASS/ISSUES)

Role-based pipelines are CrewAI's sweet spot — describe roles / goals / tasks, framework orchestrates.

How to run — two paths¶

Path A (default, free, local)¶

pip install -r requirements.txt
ollama pull qwen2.5:3b
ollama serve
python starter.py

Budget: $0. Three agents sequential ≈ 30-90s on CPU with qwen2.5:3b.

Path B (Anthropic, cloud-quality)¶

pip install -r requirements.txt
export ANTHROPIC_API_KEY=sk-ant-...
python starter_anthropic.py

Budget: ~$0.005-0.01 per run (3 agents × short outputs, claude-haiku-4-5).

Validate the logic¶

python test.py             # tool logic + crew structure
python test_anthropic.py   # starter_anthropic loads

CrewAI's kickoff() is too opaque for pure mock testing. These tests cover structure (3 agents + 3 tasks + sequential process + context dependencies) and tool logic. For full validation, run starter.py against Ollama.

Core CrewAI multi-agent concepts¶

Agent¶

researcher = Agent(
    role="Researcher",
    goal="...",          # one line: what does "success" look like
    backstory="...",     # persona context, shapes the prompt
    tools=[search],
    llm=MODEL,
)

Key: role and goal dramatically affect prompt quality. Don't write "Agent" — write "Researcher who finds factual data".

Task¶

research_task = Task(
    description="Search for X and report findings.",
    expected_output="A 1-2 sentence factual entry.",
    agent=researcher,
)

Key: expected_output is the "passing template" the LLM sees — be specific. "A 2-sentence intro paragraph" is 10× better than "Some text".

Context dependency¶

write_task = Task(..., context=[research_task])   # writer sees researcher's output
critic_task = Task(..., context=[research_task, write_task])  # critic sees both

Key: context is CrewAI's dataflow mechanism. critic_task.context=[a, b] means the critic sees the output of tasks a and b.

Sequential vs hierarchical process¶

Crew(..., process=Process.sequential)    # linear walk-through
Crew(..., process=Process.hierarchical)  # manager + workers, needs manager_llm

We use sequential (simplest, deterministic). Hierarchical lets a manager agent dispatch — useful for more complex tasks.

Observation across both paths¶

Observation	Anthropic Claude haiku	Ollama qwen2.5:3b
Researcher actually calls tool	Stable	Sometimes skips tool, makes things up
Writer cites researcher's output	Stable	May go off-memory, deviate from search result
Critic catches hallucinations	Sharper	Looser, may PASS when it shouldn't
Speed	10-30s	30-90s
Cost	$0.005-0.01	$0

Punchline: multi-agent is more model-quality-sensitive than single-agent — each agent can drop a step, errors compound by the time you reach the critic. Production multi-agent systems almost always use large models (or carefully fine-tuned small ones).

Common pitfalls¶

expected_output too generic: "Some output" gives the LLM no guide. "A 2-sentence blog intro paragraph in active voice" is 10× better
Missing context: writer without context=[research_task] doesn't see researcher's output — it'll hallucinate
Small model + 3 agents: qwen2.5:3b on 3-agent crews can take 1+ minute. Switch to qwen2.5:7b or Claude
allow_delegation=True use cautiously: enables agents to call others, easy to loop. Default False for prototypes

Want smarter answers?¶

MODEL=anthropic/claude-sonnet-4-6 python starter_anthropic.py  # higher quality
MODEL=ollama/qwen2.5:7b python starter.py                       # larger local model

Extensions¶

Add a manager: process=Process.hierarchical + manager_llm=... for dynamic delegation
Add memory: CrewAI has memory=True for cross-task context
Streaming: crew.kickoff_for_each(...) or crew.kickoff_async(...)
Human-in-the-loop: see Exercise 3 (LangGraph) — CrewAI's HITL is weaker