CREWAI

CrewAI Python crews role-based agents.

Définition

CrewAI key concepts : (1) Agent : defined by role (e.g., 'Senior Research Analyst'), goal (e.g., 'find latest AI news'), backstory (LLM character context), LLM (any LangChain LLM compatible), tools (LangChain tools or custom). (2) Task : assigned to specific agent, description (what agent should do), expected_output (what should produce), tools_required. (3) Crew : collection of agents + tasks, process type (sequential = tasks executed in order, hierarchical = manager LLM coordinates other agents). (4) Memory : ShortTermMemory (current task), LongTermMemory (across tasks in crew), EntityMemory (entities mentioned). (5) Process types : (a) Sequential : tasks executed in defined order, each agent works on their task with potentially results from previous tasks as context. (b) Hierarchical : crew has a manager agent that delegates tasks, automatic coordination, more flexible. (6) Custom Tools : Tool class wraps Python functions, integrates with LangChain tools ecosystem (Tavily Web Search, Wikipedia, file reading, etc.). CrewAI v0.x focus iteration on agent role-clarity (vs free-form agent conversations AutoGen). Customers : ~50000+ deployments 2024, focus startups + content marketing + research workflows automation.

Origine

CrewAI initial release septembre 2023 par Joao Moura ; CrewAI Inc. fondee 2024 ; Seed $18M 2024 (Insight Partners + Boldstart Ventures) ; ~20000+ GitHub stars 2024.

Exemple en contexte

Content marketing crew CrewAI : agents = ResearchAnalyst (researches trending topics Tavily Web Search), ContentWriter (writes draft article based research), Editor (edits + polishes draft) ; tasks sequential : 1) ResearchAnalyst finds trending AI agentic frameworks topic, 2) ContentWriter writes 1500-word draft, 3) Editor polishes ; output : publishable blog article in ~5 minutes via OpenAI GPT-4.

Termes liés

• AutoGen — alternative multi-agent.