CrewAI

CrewAI is an open-source Python framework for building multi-agent systems, in which teams of role-based agents, called crews, work on tasks, and event-driven Flows control the surrounding application logic.

Core concepts. An agent has a role, a goal and tools. A task is a unit of work with an expected output. A crew groups agents and tasks, and a process decides how the work runs. Flows manage state and event-driven execution, with conditional branches and loops, and hand complex tasks to crews, which return their results to the Flow. The framework is MIT-licensed. The latest release on GitHub at the time of writing is 1.15.22, from September 16, 2026.

Processes. In the sequential process, tasks run in a set order and each output feeds the next. In the hierarchical process, a manager agent, configured with manager_llm or manager_agent, plans the work, assigns tasks to agents according to their capabilities, and reviews the results. Tasks are not pre-assigned.

Planning and memory. With planning=True, an AgentPlanner runs before each crew iteration and adds a step-by-step plan to every task description. Memory is a single Memory class: a model infers the scope and importance of each item when saving, and recall blends semantic similarity, recency and importance.

A2A. CrewAI’s documentation treats A2A as a delegation primitive. With A2AClientConfig, an agent decides per task whether to handle it locally or delegate to a remote A2A agent. With A2AServerConfig, a CrewAI agent is exposed as an A2A server with a generated Agent Card. The documented transports are JSON-RPC, gRPC and HTTP+JSON, and the server configuration’s protocol_version defaults to 0.3.0, so check what your counterpart expects when pairing it with an A2A v1.0 agent. A2A support is installed with the crewai[a2a] extra, which uses the official a2a-sdk package.

Neighbouring terms. An agent framework is the general category. The hierarchical manager is CrewAI’s orchestrator agent, and the AgentPlanner is its planner.