CrewAI: Role-Based Multi-Agent Orchestration#
Open-source Python framework by João Moura for role-based multi-agent systems. Agents defined by role + goal + backstory, organized into “crews” with sequential or hierarchical process strategies.
Four Core Abstractions#
- Agent — role, goal, backstory, tools. Backstory shapes personality.
- Task — work assigned to agent, with dependencies via
contextparameter. - Crew — collection of agents + tasks + process strategy.
- Process — sequential (ordered) or hierarchical (manager delegates).
Memory System#
- Short-term — context within current crew execution
- Long-term — lessons learned across executions
- Entity memory — info about specific entities encountered
Cross-agent knowledge sharing and stateful execution built in.
Key Design: Role + Backstory#
Unlike other frameworks where agents are just system prompts, CrewAI’s backstory shapes the agent’s personality and approach. “Senior Security Auditor” with a detailed backstory produces different output than a generic “Reviewer.” This is a form of persona-based context-management.