01 · First Crew

Your first multi-agent workflow: researcher + reporter running sequentially.

Prerequisites

• Completed: 00 · Installation
• ollama list shows qwen3:8b

Time: 5 minutes End state: You have run a main.py locally and seen a two-stage Markdown report as output.

1. Create the project

mkdir my_first_crew && cd my_first_crew

Create main.py with the following content (from spec §1.4 — the PKV example, simplified to smart-home research):

"""First Crew — sequential researcher+reporter pattern."""

from __future__ import annotations

import asyncio

from cognithor.crew import Crew, CrewAgent, CrewProcess, CrewTask


def build_crew() -> Crew:
    researcher = CrewAgent(
        role="Researcher",
        goal="Research facts on the given topic",
        llm="ollama/qwen3:8b",
    )
    reporter = CrewAgent(
        role="Reporter",
        goal="Write a structured report",
        llm="ollama/qwen3:8b",
    )
    research = CrewTask(
        description="Research: Home-automation trends in 2026",
        expected_output="Bullet points of the 5 most important trends",
        agent=researcher,
    )
    report = CrewTask(
        description="Produce a report based on the research",
        expected_output="Markdown report, 300 words",
        agent=reporter,
        context=[research],
    )
    return Crew(
        agents=[researcher, reporter],
        tasks=[research, report],
        process=CrewProcess.SEQUENTIAL,
    )


def main() -> None:
    crew = build_crew()
    result = asyncio.run(crew.kickoff_async())
    print(result.raw)


if __name__ == "__main__":
    main()

2. Run it

python main.py

The first run can take 30–60 s while Ollama warms up the model. Subsequent runs are much faster.

Expected output (truncated):

# Home-Automation Trends 2026

- Matter/Thread adoption on >60% of new devices
- Local AI inference (edge LLMs) for voice control
- ...

3. What happened?

1. CrewAgent declares only the role — no prompts, no code.
2. CrewTask declares the work, plus dependencies via context=[research].
3. Crew(process=SEQUENTIAL) compiles both tasks into a DAG and runs them in order.
4. Every task internally goes through the PGE-Trinity (Planner → Gatekeeper → Executor), exactly like any other Cognithor action. No new security surface.
5. The final result.raw is the output of the last task. All intermediate outputs are in result.tasks_output.

4. Runnable copy in the repo

The same example lives as a standalone mini-project at examples/quickstart/01_first_crew/. You'll find:

• main.py — identical script
• requirements.txt — just cognithor>=0.93.0
• test_example.py — a smoke test with mocked planner, runs in CI

5. Variations

• More agents: add any number of CrewAgent instances to agents=.
• Hierarchical: set process=CrewProcess.HIERARCHICAL + manager=CrewAgent(...) to let a manager agent delegate tasks.
• YAML config: instead of Python, use agents.yaml + tasks.yaml — see cognithor.crew.yaml_loader.

Next: 02 · First Tool