Orchestrate Multi-Agent AI Crews with CrewAI

This skill takes on the persona of a CrewAI Multi-Agent Architect, positioning CrewAI as the leading role-based multi-agent framework - the source describes it as used by 60% of Fortune 500 companies. It covers agent design (role, goal, backstory), task decomposition with dependencies, crew orchestration, process selection (sequential or hierarchical, with parallel also named as a process type), memory configuration, and Flows for event-driven multi-stage workflows. The recommended pattern defines agents and tasks in YAML config files and wires them together in a @CrewBase class:

from crewai import Agent, Task, Crew, Process
from crewai.project import CrewBase, agent, task, crew

@CrewBase
class ContentCrew:
    agents_config = 'config/agents.yaml'
    tasks_config = 'config/tasks.yaml'

    @agent
    def researcher(self) -> Agent:
        return Agent(config=self.agents_config['researcher'])

    @crew
    def crew(self) -> Crew:
        return Crew(
            agents=self.agents,
            tasks=self.tasks,
            process=Process.sequential,
            verbose=True
        )

Use it for structured, role-based multi-agent workflows in Python - sequential pipelines, hierarchical crews where a manager agent (via manager_llm) delegates to specialized workers, or event-driven multi-stage Flows that persist state across steps with @start/@listen/@router decorators. It's Python-only, can be verbose for simple cases, and Flows are a newer feature; the skill's own delegation triggers point elsewhere when the actual need is different - LangGraph for explicit state-machine/graph management, Langfuse for LLM observability and tracing, or structured-output for JSON-schema-constrained responses.

Requires Python 3.10+, the crewai package, and LLM API access. A crew's kickoff(inputs={...}) call runs the defined process and returns a result; enabling planning=True with a planning_llm makes CrewAI generate a step-by-step plan first and inject it into each task before execution, accessible afterward via crew.plan. Memory is enabled with memory=True for short-term (within a task), long-term (across executions), and entity memory, or customized with explicit LongTermMemory/ShortTermMemory objects, a custom storage backend, and an embedder configuration. Custom tools are added either as a BaseTool subclass with a Pydantic args_schema and a _run method, or via the @tool("name") function decorator, then passed to an agent's tools list.

Built on the CrewAI framework and CrewAI Tools, with common integrations for OpenAI, Anthropic, and Ollama as LLM providers, SerperDevTool for web search, FileReadTool/DirectoryReadTool for local file access, and custom tools; deployed as Python applications, FastAPI backends, or enterprise systems. Related skills it composes with: langgraph, langfuse, structured-output, and autonomous-agents - documented composition patterns include a research-then-writing pipeline chained via task context and structured output, an observable agent team with a Langfuse callback handler monitoring interactions, and a complex Flow that combines LangGraph-style state handling with CrewAI crews for branching and routing.

Developers building collaborative AI agent teams in Python who need clear role/goal/backstory agent design, dependency-aware task decomposition, and a choice between sequential and manager-led hierarchical orchestration - with memory and Flows available once workflows get more complex.