CrewAI — The Complete Guide
Instead of a single agent, a team of specialists working together
CrewAI is an open-source Python orchestration framework by João Moura / CrewAI Inc., not just a library — a full multi-agent platform with Crews, Tasks, Agents, Flows (event-driven workflows added in 2024), and built-in tools (SerperDev, WebsiteSearchTool, ScrapeTool and more). It supports 100+ LLM providers through LiteLLM: Anthropic (Claude Sonnet 4.6, Opus 4.7, Haiku 4.5), Google Gemini 2.5 Pro/Flash, OpenAI, Groq, DeepSeek V3, Mistral, and local models via Ollama. Each Agent is defined with a role, a goal, a set of tools, and its own LLM; workflows run as `sequential` or `hierarchical` Processes, or as event-driven Flows. A typical deployment runs behind FastAPI + Docker. I currently run 10 crews on my VPS (blog-he, marketing-team, yt-to-blog-he, research-crew and more) — but for you, CrewAI can power content automation, research ops, distributed code review, data analysis, customer research, or anything that needs more than a single prompt to a single LLM.
What this guide covers
What is CrewAI? An AI team that works for you
An open-source Python orchestration framework for multi-agent workflows with Flows, Crews, and 100+ LLM providers via LiteLLM
CrewAI is a Python framework by João Moura and CrewAI Inc. that began as an open-source project in 2023, born from a simple realization: no matter how capable a single language model is, it rarely produces the best result on genuinely complex tasks. In 2026 it is one of the most mature multi-agent platforms available, with support for 100+ LLM providers through LiteLLM (Anthropic, Google, OpenAI, Groq, DeepSeek, Mistral, Ollama and more), Flows for event-driven pipelines (added in 2024), Crews for collaborative teams, and a rich built-in tool set (SerperDev, WebsiteSearchTool, FileReadTool and more). CrewAI's approach is simple but elegant: instead of asking one AI to write an entire article, you define a team of agents — each with a clear role, a defined goal, and access to its own tools — and let them work together, the way a human team would. The result is deeper, more accurate, and usually cheaper, because you can reserve your strongest model for the parts that truly need it. An Agent is the basic worker; a Task is a unit of work; a Crew is the whole team plus the Process that ties them together; and an LLM is the model that powers each agent.
Agent, Task, Crew — the core concepts
Four ideas that cover 90% of what you'll ever need
CrewAI's structure feels like running a virtual team: you have agents (the workers), tasks (what needs doing), and the crew itself (how they operate together). Every concept here is a Lego brick you can snap together — once you internalize the four fundamentals, any workflow you can imagine becomes reachable. The code is clean enough that my small crews come in under 30 lines of Python, and even the larger ones powering my full projects rarely exceed 100 lines.
Your first crew in 15 minutes
Install, configure, and run
Your first crew is the classic research → writing → editing pipeline, exactly like the way a real content team splits its work. You provide a topic, and your virtual team returns a polished article. It is no more than 30 lines of Python, and anyone coming from Claude Code or the agent world will recognize the philosophy immediately. You can run this crew locally on your machine or, like mine, host it on a VPS and call it through an API.
3 real crews in production
Education pipeline, marketing team, YT → blog
Theory is great — but here is how it looks in the wild. These are three crews I run daily on my VPS, producing real content. Each one is a role split that mirrors a human team — someone researches, someone shapes the message, someone writes in fluent Hebrew. You can clone the skeleton and adapt it to your domain: a legal advisory crew, a customer support crew, a data analysis crew — same principles, different roles.
Gemini + CrewAI — what to know up front
ReAct parser issues on long inputs
This is the trap that cost me a few sleepless nights, so it is worth flagging early. CrewAI orchestrates its agents using a universal format called ReAct (Reasoning + Action) — a kind of internal language between the team and its conductor. Gemini, smart as it is, does not always follow that format precisely when inputs get long. The result: retry loops, timeouts, and a stuck crew. My yt_to_blog_he crew hit this exact wall on long video transcripts, and the fix required some creative engineering.
Hierarchical processes + a manager agent
When sequential isn't enough
Process.hierarchical is where CrewAI stops being a relay race and becomes a real orchestra with a conductor. Instead of tasks moving in a straight line, a manager agent (typically a strong model like Claude Opus 4.7 or GPT-5) makes real-time decisions: who gets which task, when to split, when to combine, and when to stop. It is advanced territory — only one of my ten crews runs this way — everything else is sequential, simply because it is more predictable and cheaper. But when you genuinely need flexibility, hierarchical is worth the effort.
