crewai

Core Concepts

• Agents: Role-based autonomous entities with role, goal, backstory, llm_config
• Tasks: Atomic units of work assigned to agents with description, expected_output, agent
• Tools: Functions agents can call (@tool decorator or Tool.from_function) for external actions
• Process Flow: Process.sequential (chain) or Process.hierarchical (manager-agent routing)
• Memory: Short-term, long-term, and entity memory between crew runs
• Custom Tools: Extend base Tool class or use Tool.from_function with input schema

Common Patterns

• Agent(role="Researcher", goal="Find information", backstory="...", llm=llm_config)
• Task(description="Analyze trends", expected_output="Report", agent=researcher)
• Crew(agents=[researcher, writer], tasks=[research, write], process=Process.sequential).kickoff()
• @tool("search_tool") decorator on functions agents can invoke
• Tool.from_function(name="calculator", func=calc, description="Adds numbers")
• Custom LLM: config={"provider": "openai", "model": "gpt-4", "temperature": 0.7}

Best Practices

• Keep agent goal specific and backstory concise for consistent behavior
• Define expected_output clearly on every task to guide agent completion
• Use Process.hierarchical only when tasks require delegation decisions
• Enable memory=True on Crew for context persistence across iterations
• Pin crewai and crewai-tools versions for reproducible agent behavior