CrewAI

CrewAI is an open-source Python framework for building multi-agent applications with its own orchestration primitives. Its two main abstractions address different parts of an automation: Crews group specialized agents around roles, goals and tasks, while Flows define event-driven logic, shared state, branches and handoffs. A Flow can call Python functions, individual agents or complete Crews, keeping autonomous collaboration inside an explicit application workflow.

The core CrewAI repository is released under the MIT license. CrewAI also offers the separate commercial AMP suite, which adds a managed control plane, deployment, observability, governance, security features and enterprise support; those hosted capabilities should not be confused with the license or feature set of the open-source framework itself. The framework can connect to multiple hosted or local model providers and can be extended with custom tools and external services.

Top features

• Role-based Crews: Define agents with goals, tools and tasks, then run work sequentially or use a manager-led hierarchical process.
• Event-driven Flows: Connect Python methods, agents and Crews with start, listen and router controls, including branching, loops and parallel entry points.
• Structured state and output: Use Pydantic-backed state and return raw, JSON or typed Pydantic results from crew execution.
• Recovery and review: Persist Flow state, checkpoint crews, flows or agents, resume interrupted work and add human-feedback gates.
• Memory, knowledge and tools: Attach reusable knowledge sources, execution memory, custom actions and packaged tools to agents and crews.
• Integration surfaces: Connect model providers, expose or consume MCP tools, stream execution events and add OpenTelemetry-compatible observability integrations.
• Project tooling: Scaffold and run projects through the CLI, using current JSONC configuration or the supported classic Python/YAML pattern.

Use cases

• Assign research, analysis and report assembly to specialist agents while a Flow controls inputs and delivery.
• Build support or operations automations that route cases, call tools and request human judgment at defined decision points.
• Process documents or business data through extraction, validation, enrichment and structured-output stages.
• Coordinate multiple crews inside a larger event-driven application without hiding ordinary Python logic.
• Resume long-running agent jobs after interruptions instead of repeating completed tasks.