Workflows/Set up a local AI agent team

Agents / WORKFLOW

Set up a local AI agent team

Run a two-agent planner and reviewer on one machine using Ollama, Qwen3 8B, and CrewAI. Verify their sequential handoff before adding tools or private data.

EstimatedSource checked 9/25/2026

LOCALRENTED GPU

Before you begin

Software: Ollama, CrewAI, Python

Hardware

One Qwen3 model stays loaded while two agents take turns. The tiers below are memory estimates for a short sequential task, not throughput measurements.

Machine	Qwen3 8B · two agents
NVIDIA GPU · 8 GB	May require CPU offload; slow
NVIDIA GPU · 10–12 GB	Minimum starting range · estimated
NVIDIA GPU · 16 GB+	Recommended starting range · estimated
RTX 5090 · 32 GB	Comfortable capacity for this short task
Apple Silicon · 16 GB	Minimum starting range · estimated
Apple Silicon · 24 GB+	Recommended starting range · estimated
CPU-only · 32 GB RAM	Can run sequentially; expect slower responses

Longer agent histories or parallel teams need more memory. This guide uses sequential agents on one machine, not a distributed cluster.

AGENTS / LOCAL TEAMSOURCE-LINKED SETUP

A real two-agent handoff on your own machine.

Build one planner and one reviewer in CrewAI, both served by local Qwen3 through Ollama. This first setup intentionally grants no external tools.

Choose a package Run its commands Check the result

BEFORE YOU BEGIN

The documented package

This setup uses the package documented for this task. Review its source and supported platforms before starting.

Save your machine in My Hardware to get an automatic starting choice. You can always choose any package yourself.

CURRENT SETUP

Qwen3 8B · Ollama · official

These are estimated starting points for a short two-agent run. The 5.2 GB package alone does not include runtime, context, and Python overhead.

View weight source

THE WORKFLOW

Set up Qwen3 8B · Ollama on your machine

Pick your operating system. Every command below is for the selected package and runtime.

Use Terminal for the commands below. This creates a two-agent sequential team on one machine; it is not a multi-node GPU cluster.

Install the local runtime on Mac

Download and open the Ollama macOS app. Install Python 3.10–3.13 and uv, then open Terminal.

Download Ollama Install uv CrewAI Python requirements

Check Python, uv, and Ollama

In Terminal, verify the supported Python version and local tools before creating the project.

python3 --version
uv --version
ollama --version

Download the local model

Run the official Qwen3 8B package once to download its 5.2 GB weights and confirm a reply. Leave Ollama running; exit the chat before the next step.

Qwen3 8B package

ollama run qwen3:8b

Create an isolated CrewAI project

In Terminal, make a clean Python project and add CrewAI’s LiteLLM extra, which its documentation requires for the Ollama provider.

CrewAI Ollama provider

uv init local-agent-team
cd local-agent-team
uv add "crewai[litellm]"

Define planner and reviewer agents

Save this as team.py in the project folder. Both agents point to the local Ollama endpoint; the reviewer receives the planner’s output through a sequential CrewAI process. No file, shell, or network tools are granted.

CrewAI agents and crews

from crewai import Agent, Crew, LLM, Process, Task

llm = LLM(model="ollama/qwen3:8b", base_url="http://localhost:11434")
planner = Agent(role="Planner", goal="Create a clear three-step plan", backstory="You break small tasks into verifiable steps.", llm=llm, verbose=True)
reviewer = Agent(role="Reviewer", goal="Find errors and improve clarity", backstory="You check each claim against the task and the plan.", llm=llm, verbose=True)
plan = Task(description="Plan a short checklist for organizing a local photo library. Do not assume access to the user's files.", expected_output="Three specific, safe steps.", agent=planner)
review = Task(description="Review the planner's checklist. Flag assumptions and produce a corrected final checklist.", expected_output="A corrected checklist with one caution.", agent=reviewer)
crew = Crew(agents=[planner, reviewer], tasks=[plan, review], process=Process.sequential, verbose=True)
result = crew.kickoff()
print(result.raw)

Run and inspect the handoff

From the project folder in Terminal, run the script. Watch the planner output, then the reviewer output, and check that the final checklist includes a correction and caution.

uv run python team.py

SUCCESS CHECK · Qwen3 8B · Ollama

Planner → reviewer → final checklist

The terminal shows a planner response followed by a reviewer response, and the final checklist reflects the reviewer’s correction. Confirm Ollama serves the requests locally.

This source-linked example combines the official CrewAI and Ollama APIs. YouRunAI has not physically tested the exact software versions on every platform.

Back to setup steps

When it doesn’t go to plan

CrewAI tries to reach a remote provider

Confirm both Agent objects use the same LLM configured as ollama/qwen3:8b with base_url http://localhost:11434. The example provides no external tools or remote API keys.

The reviewer does not receive a useful plan

Confirm Process.sequential and the task order, then ask for a shorter checklist. Inspect the verbose output for each agent before adding more roles.

The model stalls or runs out of memory

Close other model sessions, keep this first task short, and select a smaller documented Ollama package only after updating the model name in both the download command and team.py.

REFERENCE LIBRARY

Sources and files

Original instructions, model files, and compatibility notes behind this setup.

4 SOURCES

01Official Qwen3 8B Ollama packageollama.com 02CrewAI installation and Python versiondocs.crewai.com 03CrewAI Ollama provider configurationdocs.crewai.com 04CrewAI crew and task APIdocs.crewai.com

The model behind this workflow

Qwen3 8B

MY HARDWARE

Will it run on your machine?

Save your machine to see a personalized rating and its reasoning.

Add my hardware

Keep exploring

Text

SETUP GUIDE

Set up a private AI chat assistant

Install a local runtime, run Qwen3.5 9B, confirm responses, and know when to choose the smaller 4B package.

Build a private local coding assistant

Connect an open coding-capable model in Ollama to Cline, run a small repository task, and review the result locally.

SOFTWAREOllama + Cline

Generate images on your own GPU

Use the official FLUX.2 Klein 4B ComfyUI template with exact model files, a first prompt, and an output check.

SOFTWAREComfyUI

LOCALRENTED GPU

Verified source