CrewAI is an open-source Python framework for building "crews" of agents that work together to accomplish a task. Each agent has a role, a goal, and tools. You orchestrate the crew, and the agents do the work.

Why it matters for non-engineers: CrewAI lets you wire together multi-step business workflows that would otherwise require hiring an employee or paying a SaaS vendor.

Examples:

A roofer's customer-intake crew:

• Agent 1 (Receptionist): reads incoming emails and texts, classifies them as "new lead," "existing customer," "vendor," or "other."
• Agent 2 (Scheduler): for new leads, checks the calendar, finds open estimate slots, drafts a reply with three options.
• Agent 3 (Follow-up): for customers whose appointments are tomorrow, sends a reminder text the day before.

A restaurant manager's daily prep crew:

• Agent 1 (Sales analyst): reads yesterday's POS data, identifies the top 10 sellers and the bottom 5.
• Agent 2 (Inventory checker): cross-references current inventory against the top sellers, flags items running low.
• Agent 3 (Order drafter): drafts an order to the food vendor based on the flagged items, sends to the manager for approval.

A consultant's research crew:

• Agent 1 (News scanner): pulls today's headlines on three topics the consultant tracks.
• Agent 2 (Summarizer): writes a 200-word brief on each topic.
• Agent 3 (Distributor): emails the briefs to the consultant's morning email at 6:30 AM.

Each of these crews is 50-200 lines of Python. We'll see real ones in the templates (sections 8.7, 8.8, 8.9).

Install: pip install crewai crewai-tools.

The mental model: each agent is a job description ("You are a customer intake specialist. Your job is to read incoming emails and classify them. You have access to the Gmail tool."). The crew is the org chart. You define the agents and the crew, and the framework handles the coordination.

CrewAI works with frontier models (Claude, GPT, Gemini) and with self-hosted models (anything that speaks the OpenAI-compatible API, which Ollama and vLLM both do). For agents running 24/7 on routine tasks, self-hosted is usually the right call (cost, privacy, latency). For agents handling complex one-off reasoning, frontier models are worth the spend.

We'll cover the small-business agent stack (Mac mini + Ollama + CrewAI) in section 4.4 and the deployment in 6.6.

================================================================================ PART 4 — MODELS, COMPUTE, AND COST