What is CrewAI

Open-source Python framework for multi-agent orchestration — role-based 'Crews' for autonomous collaboration and event-driven 'Flows' for precise control, now a de facto standard for production agentic pipelines.

CrewAI is an open-source Python framework for building multi-agent systems — several LLM agents collaborating on one task instead of a single agent doing everything. It’s become one of the most widely adopted frameworks in this space, built around a metaphor that maps directly onto how humans organize teams.

The core metaphor

Every agent gets three things, the same way you’d brief a new hire:

Field	Purpose
Role	What this agent is (“Senior Data Analyst”, “QA Reviewer”)
Goal	What it’s trying to achieve on this task
Backstory	Context that shapes tone, priorities, and judgment calls

Agents are assigned Tasks, and Tasks are grouped into a Crew that executes them under one of three process types:

Sequential    Agent A → Agent B → Agent C            (fixed pipeline)
Hierarchical  Manager agent ──▶ delegates to workers  (dynamic delegation)
Consensual    Agents vote on a decision before proceeding

Crews vs. Flows

This is the distinction that matters most when picking CrewAI for a real system:

Crews — autonomous agent collaboration. You describe roles and goals; the agents figure out the “how.” Good for open-ended reasoning tasks.
Flows — event-driven, low-level control. You define exact state transitions and routing logic, calling into Crews as steps when you need agentic reasoning inside an otherwise deterministic pipeline.

Flow (deterministic backbone)
  ├── Step 1: fetch data (plain code)
  ├── Step 2: Crew("triage") — agents reason about the data
  ├── Step 3: if triage.severity == "high": Crew("investigate")
  └── Step 4: write result (plain code)

Production systems tend to use Flows as the skeleton and Crews only where genuine reasoning is needed — this avoids the classic failure mode of letting agents “decide” things a simple if statement should have decided.

Recent architecture additions (2026)

Pluggable backends for memory, knowledge, and RAG — swap in What is Mem0 or another memory provider instead of the default, without rewriting the Crew.
Chat API for conversational (multi-turn) flows, not just one-shot task execution.
Scoped runtime state — isolates state between concurrent runs so parallel Crew executions don’t leak context into each other. This matters the moment you run CrewAI as a shared service rather than a single local script.
Native LLM providers beyond OpenAI/Anthropic, including Snowflake Cortex.

Where it fits next to the rest of the agent stack

Concern	CrewAI’s answer
Orchestration	Native — Crews (autonomous) + Flows (deterministic)
Memory	Pluggable — bring your own, e.g. What is Mem0
Tool access to your data	Bring your own MCP client, or wrap What is MCP Toolbox as a CrewAI tool
Browser/computer control	Not native — pair with What is Playwright MCP as a tool
Deployment	Your own process/container — no managed runtime like What is Vertex AI‘s Agent Engine

Why it’s on the backlog: it’s the most direct comparison point for What is Google ADK — both solve multi-agent orchestration, but ADK is Google’s “agent execution framework” tied to its own runtime and languages, while CrewAI is runtime-agnostic Python you can drop into any pipeline, including an h-aiops SRE-agent step.