crewai-multi-agent

Define Agent objects with role, goal, and backstory, create Task objects assigned to each agent, then combine them in a Crew with Process.sequential or Process.hierarchical. Call crew.kickoff with your inputs to execute the workflow.

CrewAI uses role-based agents with a lower learning curve and built-in memory, suited for team-style task delegation. LangGraph uses graph-based state machines, better for complex stateful workflows with cycles. CrewAI has no LangChain dependency.

Set the context parameter on the downstream task with a list of previous Task objects, for example context=[research_task]. The receiving agent then gets the earlier task's output as input when it executes.

Yes, CrewAI supports local models through the LLM class, for example LLM(model="ollama/llama3.1", base_url="http://localhost:11434"). It also works with OpenAI, Anthropic, and Azure endpoints.

Agents loop when tasks are ambiguous or iterations are unlimited. Set max_iter on the agent to cap reasoning steps, add max_rpm for rate limiting, and write task descriptions with explicit expected outputs such as an exact number of bullet points.

Use Flows when you need conditional branching, complex state management, or event-driven execution with routers and listeners. Crews alone are sufficient for straightforward sequential or hierarchical multi-agent collaboration.