Week 1 — An agent is a job description¶

Course: CrewAI
Who this is for: Engineers who have written a role into a hiring doc and an IAM policy.

CrewAI is jobs + tickets + a sprint team. An agent is a worker with a system prompt (role, goal, backstory), tools (IAM), and a task (a ticket). It is ceremony when one function would do.

This environment pins crewai==0.80.0 with no LangChain and no crewai-tools. Concept demos construct objects and assert fields. They do not kickoff() a ReAct loop against a fake LangChain LLM.

🎯 What you will be able to do¶

Construct Agent(role, goal, backstory) and assert the fields
Write a tool as a local dict lookup (no Serper, no network)
Know when a second agent is staffing theatre
Leave Crew.kickoff() for a run that has a real model — not this demo

Think of it like… a hiring packet, not a hive mind.

Role = job title. Goal = OKR. Backstory = the prompt’s few-shot personality. Tools = IAM. If you cannot say what output contract the worker owes, you are not ready to add a second worker.

Picture one worker¶

Agent
  role:      "CloudWave ticket analyst"
  goal:      "Extract category and whether to escalate"
  backstory: "You read tickets. You do not invent MRR."
  tools:     lookup_plan(user_id) → local dict
                │
                ▼
         (later) Task + Crew
                │
                ▼
         kickoff()  ← needs a real model; not this week’s demo

Hypothetical CloudWave: a few ticket analysts. Not a HubSpot/Jasper case study.

Construct and assert — no kickoff¶

from crewai import Agent

analyst = Agent(
    role="CloudWave ticket analyst",
    goal="Extract category (bug, billing, question) and escalate: bool",
    backstory="You read support tickets. You do not invent metrics or refunds.",
    verbose=False,
    allow_delegation=False,
)

assert analyst.role == "CloudWave ticket analyst"
assert "escalate" in analyst.goal.lower()
assert "do not invent" in analyst.backstory.lower()
assert analyst.allow_delegation is False

If construction tries to bind a default OpenAI client, set a dummy OPENAI_API_KEY in the shell for the constructor only — still do not kickoff(). The fields are what this week proves.

A Crew without a run is a roster:

from crewai import Crew, Process, Task

task = Task(
    description="Classify this CloudWave ticket: '{ticket}'",
    expected_output="JSON with keys category, escalate",
    agent=analyst,
)
crew = Crew(agents=[analyst], tasks=[task], process=Process.sequential)
assert crew.agents[0].role == analyst.role
assert crew.tasks[0].expected_output.startswith("JSON")
# do not crew.kickoff() in this concept demo

A tool is a local lookup¶

No crewai_tools, no SerperDevTool, no LangChain tool wrapper.

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


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


assert lookup_plan("user_0001") == "Enterprise"
assert lookup_plan("nope") == "unknown"

Library spelling if your 0.80 install exposes it (optional):

from crewai.tools import tool

@tool("lookup_plan")
def lookup_plan_tool(user_id: str) -> str:
    """Look up a CloudWave plan. Local dict only."""
    return PLANS.get(user_id, "unknown")

Attach tools only if you can do it without importing langchain or crewai_tools. Otherwise keep the function and assert it — that is the IAM idea.

Watch out — kickoff is a ReAct loop

crew.kickoff() will prompt a model, possibly call tools, possibly loop. A FakeListLLM from LangChain is out of this venv. Do not !pip install in a notebook. Do not add a second agent to “improve quality” before you have an output contract.

Ship / don’t ship

Ship one agent whose role/goal/backstory you can assert and whose tool is a function you wrote. Don’t ship a three-agent blog factory that kickoff()s against a fake LLM, and don’t cite vendor marketing as evidence.

What this week is not¶

Not multi-agent consensus (week 3).
Not a 20-task launch plan (week 2 is three tasks).
Not production (week 4 compares one worker vs a crew on calls and latency).

✍️ Exercise¶

Exercises.

🤔 Reflection¶

Which CloudWave job actually needs a second role, not a second prompt?
What does lookup_plan return for an unknown user, and why is that better than raising?
If kickoff() is off-limits without a provider, what did you still prove?

🔗 Next week¶

Tasks with context=[upstream]. Three tickets, not twenty.