What Is CrewAI? Multi-Agent AI Framework Explained

Core Architecture

CrewAI organizes multi-agent systems into four layers, each serving a distinct function in the execution pipeline. Understanding this hierarchy is the prerequisite for productive use of the framework.

Flows are the top-level event-driven backbone. A Flow defines the overall sequence of operations, routes data between steps, and handles conditional branching. Think of it as the orchestration layer that decides which Crew runs when, and what happens with the results.

Crews are collaborative agent groups assigned to a specific objective. Each Crew contains one or more Agents, a set of Tasks, and a process model that determines execution order. A Crew operates as a self-contained unit: it receives inputs from a Flow, coordinates its agents, and returns structured outputs.

Agents are the individual autonomous units in this agentic AI framework. Each agent gets a role (e.g., "Senior Data Analyst"), a goal, and a backstory that shapes its behavior. Agents can be assigned specific LLMs, tools, and memory configurations independently of other agents in the same Crew.

Tasks define the specific assignments with guardrails. Each task specifies an expected output format, the agent responsible, and optional validation criteria. Tasks are where you control quality: you can require structured JSON output, set token limits, or chain task outputs as inputs to subsequent tasks.

How CrewAI Works

CrewAI's evolution from a weekend project to a funded platform follows a clear trajectory. Each milestone added a distinct capability layer to the framework.

The practical workflow follows a predictable pattern. You define agents with roles and tools, assign them tasks with expected outputs, group them into a Crew with a process model, and optionally embed that Crew inside a Flow for multi-step orchestration. Execution logs show each agent's reasoning chain, tool calls, and output, which is critical for debugging.

Pricing and Plans

CrewAI uses a freemium model. The open-source framework itself is free under the MIT license. The managed platform (CrewAI Cloud) charges based on workflow executions per month. Pricing data below is sourced from crewai.com as of May 2026.

The open-source framework is separate from the managed platform. You can run CrewAI locally with your own LLM API keys at no cost beyond the LLM provider's charges. The pricing tiers above apply only to the CrewAI Cloud managed service.

Getting Started

The fastest path to a working CrewAI installation takes about 10 minutes if you already have Python 3.10+ and an LLM API key. Here is what the initial setup involves.

Install the framework with pip install crewai, then scaffold a new project with crewai create crew my-project. The CLI generates a directory structure with YAML configuration files for agents and tasks, plus a Python entry point. Edit the YAML to define your agent roles, goals, and tool assignments, then run crewai run to execute.

CrewAI connects to LLM providers through environment variables. Set your API key (e.g., OPENAI_API_KEY for OpenAI or ANTHROPIC_API_KEY for Anthropic) and the framework handles the rest. Through native support and LiteLLM integration, CrewAI works with OpenAI, Anthropic, Google, Azure, AWS Bedrock, Meta Llama, Mistral, Groq, Ollama, and more than a dozen additional providers.