CrewAI

crewAIInc/crewAI — an open-source Python framework for multi-agent orchestration. You define agents with roles, goals, and tools, then compose them two ways: Crews (role-based agents collaborating autonomously) and Flows (event-driven workflows with precise, deterministic control, single LLM calls, and native Crew support). Deliberately standalone — a lightweight core, not a LangChain wrapper — which is its main pitch versus heavier frameworks.

In the LLM App Engineering Stack this is the agent-orchestration lane: when a task needs several specialised agents (researcher → writer → reviewer) rather than one big prompt. Pairs naturally with Instructor for typed tool outputs and Langfuse for tracing (CrewAI is a first-class Langfuse integration).

🔗 Links

Description

• Repo: https://github.com/crewAIInc/crewAI
• Homepage: https://crewai.com · Docs: https://docs.crewai.com
• Learn (community courses, 100k+ certified): https://learn.crewai.com
• License: MIT

Download or use

uv pip install crewai
uv pip install 'crewai[tools]'   # + built-in tool library

from crewai import Agent, Task, Crew
 
researcher = Agent(role="Researcher", goal="Find facts", backstory="…")
task = Task(description="Research X", agent=researcher, expected_output="notes")
crew = Crew(agents=[researcher], tasks=[task])
print(crew.kickoff())

🗒️ Description

🧩 Crews vs Flows

Primitive	Optimises for	Use when
Crews	Autonomy + collaborative intelligence	Open-ended tasks where agents divide the work themselves
Flows	Precise, event-driven control	Production automations needing deterministic branching + auditability

🧩 Why it exists

• Purpose-built for agent orchestration — clean primitives (Agent / Task / Crew / Flow) instead of general-purpose graph plumbing.
• Production-leaning — event-driven Flows, a tools ecosystem, and a commercial “AMP” control plane for teams that want hosted deployment/monitoring.
• Big community footprint (100k+ developers certified through its courses).

✍️ Reasoning for

• Role decomposition — for Agentic Systems where distinct personas beat a monolithic prompt, Crews express that directly.
• Control when it matters — Flows give the deterministic, event-driven backbone that pure-autonomous frameworks lack, which is what makes agent workflows shippable rather than demo-ware.
• Standalone core keeps the dependency surface (and the LangChain lock-in) small.

Weak points: multi-agent adds cost + latency + non-determinism — only worth it when a single-agent loop genuinely can’t do the job; the AMP suite is commercial; role-based autonomy can wander without tight expected_output schemas.

Alternatives considered

• LangGraph — graph-based, lower-level control; more flexible, more plumbing.
• AutoGen (Microsoft) — conversation-centric multi-agent; research-heavy lineage.
• OpenAI Swarm / Agents SDK — minimal handoff-based orchestration, tied to OpenAI’s stack.
• Single agent + tools — the honest default; reach for CrewAI only when the task is genuinely multi-role.

🔗 Resources

• README: https://github.com/crewAIInc/crewAI
• Docs: https://docs.crewai.com
• Examples: https://github.com/crewAIInc/crewAI-examples

---

Template: tool