Skip to content

Exercises — Week 1 — One agent, one local tool

Do these after reading Week 1. crewai==1.15.22 only — no langchain, no crewai_tools, no SerperDevTool, no !pip. Do not kickoff().

from crewai import Agent

Predict before you run

Does constructing an Agent call a model? Does lookup_plan("nope") raise or return "unknown"?

Runnable command

python -c "from crewai import Agent; print(Agent.__name__)"

If CrewAI is not installed: make setup-crewai (or pip install -r requirements-crewai.txt in its own venv). Do not kickoff().

Each task has three hints, closed by default. Open only as far as you need.

1. Construct and assert

Build one CloudWave Agent(role, goal, backstory).

Checks:

  • agent.role, agent.goal, agent.backstory are the strings you passed
  • allow_delegation is False
  • You never called crew.kickoff()
Hint 1 — a nudge

An Agent is a config object — a job description. What would you expect a constructor like that to do on the network?

Hint 2 — the approach

Pass role, goal, backstory, allow_delegation=False, verbose=False, then assert each attribute back. If you want proof there's no model call, time the constructor.

Hint 3 — most of the code 
import time

from crewai import Agent, Crew, Process, Task

t0 = time.perf_counter()
agent = Agent(
    role="CloudWave plan lookup",
    goal="Answer which plan a user_id is on, from the local catalog only",
    backstory="You read the plan table. You never guess a plan.",
    allow_delegation=False,
    verbose=False,
)
print(f"constructed in {time.perf_counter() - t0:.3f}s", agent.role, agent.allow_delegation)

2. Local @tool dict lookup

Plan catalog in a dict. Implement lookup_plan(user_id) -> str. Prefer from crewai.tools import tool if it imports; otherwise a plain function is the same idea.

Checks:

  • lookup_plan("user_0001") is a known plan (Enterprise / Pro / Free)
  • Unknown id returns "unknown" (string), not an exception
  • No HTTP, no Serper, no crewai_tools import
Hint 1 — a nudge

A tool a model calls must never crash the model's turn. What should a lookup return for a key it doesn't have?

Hint 2 — the approach

Write the plain function first with dict.get(user_id, "unknown") and test it. Then wrap it with @tool("lookup_plan") — the decorator turns it into a tool object whose .name and .description come from the decorator and docstring.

Hint 3 — most of the code 
from crewai.tools import tool

PLANS = {"user_0001": "Enterprise", "user_0002": "Pro", "user_0003": "Free"}


def lookup_plan(user_id: str) -> str:
    """Return the CloudWave plan for a user_id from the local catalog."""
    return PLANS.get(user_id, "unknown")


lookup_plan_tool = tool("lookup_plan")(lookup_plan)
print(lookup_plan("user_0001"), lookup_plan("nope"), lookup_plan_tool.name)

3. Roster, not a run

Optional: wrap the agent in a Task + Crew(process=Process.sequential) and assert len(crew.agents) == 1.

Checks:

  • crew.tasks[0].agent is agent (or same role)
  • Still no kickoff()
Hint 1 — a nudge

A Crew is a roster plus a work order. What's the one method that turns it into a run?

Hint 2 — the approach

Task(description=..., expected_output=..., agent=agent), then Crew(agents=[agent], tasks=[task], process=Process.sequential). Assert on the objects; never call kickoff.

Hint 3 — most of the code 
task = Task(
    description="Which plan is {user_id} on?",
    expected_output="One word: Enterprise, Pro, Free, or unknown",
    agent=agent,
)
crew = Crew(agents=[agent], tasks=[task], process=Process.sequential)
print(len(crew.agents), crew.tasks[0].agent.role, crew.process)

Expected observation

Open after you run

agent.role is the string you passed. allow_delegation is False. Unknown user id is the string "unknown".

Self-check

No langchain, no crewai_tools, no Serper, no !pip. Roster ≠ a run.