CrewAI

CrewAI is a Python framework for building multi-agent AI systems from scratch, independent of LangChain or other agent frameworks. It offers two complementary abstractions: Crews, where role-playing autonomous agents (each with a goal, backstory, and set of tools) collaborate on tasks with configurable process orchestration (sequential or hierarchical); and Flows, an event-driven layer for precise, low-level control over execution order, state, and conditional branching between steps.

The framework owns the application’s structure once adopted — agents, tasks, tools, and crews are declared as first-class objects and the framework drives their execution, memory, and tool invocation. It ships built-in support for memory (short-term, long-term, entity), a large tool ecosystem via the separate crewai-tools package, MCP (Model Context Protocol) integration, and OpenTelemetry-based tracing. A companion CLI scaffolds new CrewAI projects.

What You Get

• Agent/Task/Crew primitives for declaring role-playing agents that collaborate via sequential or hierarchical processes
• Flow, an event-driven orchestration layer for explicit, conditional step-by-step control over multi-agent pipelines
• Built-in short-term, long-term, and entity memory plus a RAG subsystem for retrieval-augmented agent context
• MCP (Model Context Protocol) integration and a companion crewai-tools package with a broad prebuilt tool library
• A CLI (crewai create) for scaffolding new projects, plus OpenTelemetry-based tracing for observability

Common Use Cases

• Building multi-step research or content-generation pipelines where specialized agents (researcher, writer, reviewer) collaborate on one output
• Automating business processes that benefit from role-based delegation, e.g. a ‘triage agent’ routing to specialist agents
• Building agentic systems with explicit control flow via Flows when a strict sequential/conditional pipeline is required over free-form agent collaboration
• Prototyping and shipping production agent workflows without depending on LangChain’s abstractions

Under The Hood

Architecture - The GitHub monorepo splits into lib/crewai (the core framework: agent/, crew.py, process.py, flow/, memory/, rag/, events/), lib/crewai-core, lib/crewai-cli, lib/crewai-tools, and lib/crewai-files, all versioned together; a Crew composes Agent and Task objects and drives execution via a configurable Process (sequential or hierarchical), while Flow provides an independent event-driven layer for explicit step ordering, sharing the same LLM/memory/RAG substrate. Tech Stack - Python 3.10-3.13, built on Pydantic v2 for schema validation, openai/instructor for LLM calls, chromadb for vector storage/RAG, and OpenTelemetry (opentelemetry-api/-sdk/-exporter-otlp-proto-http) for tracing; dependency management and linting use uv, ruff, and mypy across the workspace. Code Quality - The tests/ tree mirrors the source layout closely (agents, crew, flow-adjacent pipeline, memory, tracing, rag, mcp, security, cli) with vcrpy/pytest-recording cassettes for deterministic LLM-call testing, bandit for security linting, and pytest-randomly/pytest-xdist for robust parallel test execution — a notably mature test setup for an AI framework. API Design - Declaring an agent (Agent(role=..., goal=..., backstory=...)) and composing a Crew(agents=[...], tasks=[...]) is intentionally close to natural language, keeping onboarding low, while Flow requires learning a separate decorator-based (@start, @listen) control-flow vocabulary for teams that need more determinism than free-form crew collaboration provides.