CtrlK
CrewAI is an open-source multi-agent orchestration framework that enables developers to build teams of AI agents that collaborate to accomplish complex tasks. Each agent is defined with a specific role, goal, and backstory, and agents are organized into "crews" that execute tasks either sequentially or hierarchically. Founded by Joao Moura in late 2023 and officially launched in January 2024, CrewAI has become one of the most widely adopted frameworks for building multi-agent AI systems, powering over 1.4 billion agentic executions and used by roughly 60% of the Fortune 500 as of late 2025 [1][2][3].
CrewAI occupies a distinctive position in the multi-agent framework landscape by prioritizing simplicity and role-based design over the lower-level flexibility offered by alternatives such as LangGraph and AutoGen. Its approach draws an analogy to human teams: rather than programming individual function calls or defining complex state graphs, developers describe agents in terms of their expertise, responsibilities, and objectives, then let the framework handle coordination and communication [4].
Joao (Joe) Moura is a Brazilian-born engineering leader with close to 20 years of experience in the software industry. He studied at New York University's Leonard N. Stern School of Business and built his career at the intersection of engineering, product design, and data science. Before founding CrewAI, Moura held roles as Lead Software Engineer at Packlane, Engineering Manager at Toptal, and Senior Engineering Manager at Clearbit, a marketing data platform that was later acquired by HubSpot. At Clearbit, Moura rose to Director of AI Engineering, where he worked extensively on AI-powered workflows [2][5].
Frustrated by the difficulty of coordinating multiple AI-powered workflows, Moura began building a framework that would allow AI agents to work together as a team. He completed the initial version in October 2023 and quietly released it as an open-source project on GitHub in November 2023 [2].
The project gained rapid traction within the developer community. Within weeks, it had accumulated thousands of GitHub stars, and by early 2024, it had become one of the fastest-growing AI projects on the platform. Moura formally launched CrewAI as a company in January 2024, serving as CEO, with veteran AI entrepreneur Rob Bailey joining as COO [2].
| Date | Milestone |
|---|---|
| November 2023 | Joao Moura releases CrewAI v0.1.0 on GitHub as an open-source project |
| January 2024 | CrewAI formally launches as a company; Moura becomes CEO |
| Mid-2024 | Framework reaches 150 enterprise beta customers within six months |
| October 2024 | CrewAI announces $18 million in total funding (inception round plus Series A); launches CrewAI Enterprise platform |
| May 2025 | Native MCP integration added to documentation and framework |
| October 2025 | CrewAI v1.0 GA released, marking production stability with locked APIs |
| November 2025 | CrewAI Signal 2025 conference held in San Francisco (sold out); India satellite event in Trivandrum; new DeepLearning.AI course launched with Andrew Ng |
| March 2026 | Framework reaches version 1.10.1 with native MCP and A2A protocol support; 45,900+ GitHub stars |
By mid-2024, the company had attracted 150 enterprise customers. In October 2024, CrewAI announced $18 million in total funding, including an inception round led by boldstart ventures and a Series A led by Insight Partners, a global software investment firm. Additional investors included Blitzscaling Ventures, Craft Ventures, and Earl Grey Capital. Notable angel investors in the round included Andrew Ng (a globally recognized AI researcher and educator) and Dharmesh Shah (co-founder and CTO of HubSpot) [1][6][7].
The funding was used to expand the team, accelerate product development, and launch CrewAI Enterprise, the company's commercial cloud platform for deploying and managing multi-agent systems at scale. In November 2025, CrewAI announced additional global expansion and partnerships, though a follow-on funding round had not been publicly disclosed as of early 2026 [1][8].
CrewAI is built around a role-based paradigm where the developer defines individual agents, assigns them to tasks, and organizes them into crews. The framework handles agent communication, task delegation, and workflow execution.
The framework is organized around five primary abstractions:
| Concept | Description |
|---|---|
| Agent | An autonomous unit with a defined role, goal, backstory, and optionally a set of tools. Agents can reason, use tools, delegate work, and communicate with other agents. |
| Task | A specific piece of work assigned to an agent, including a description, expected output, and optional context from other tasks. |
| Crew | A team of agents working together on a collection of tasks, with a defined process type and optional configuration for memory, caching, and other features. |
| Process | The execution strategy for a crew. Options include sequential (tasks executed one after another), hierarchical (a manager agent delegates to worker agents), and parallel (independent tasks run simultaneously). |
| Tool | A capability that an agent can use to interact with the external world, such as web search, file reading, API calls, or code execution. |
Agents are the building blocks of any CrewAI application. Each agent is defined with natural language descriptions of its role, goal, and backstory. These descriptions shape how the agent approaches tasks and communicates with other agents. For example, a "Senior Research Analyst" agent might have the goal of "uncovering cutting-edge developments in AI" and a backstory explaining that it is "a veteran analyst at a leading tech think tank with a passion for identifying emerging trends." This role-based approach allows developers to think about their AI workflows in terms of team composition rather than low-level programming logic [4].
Agents can be configured with various parameters including which large language model to use, whether they are allowed to delegate tasks to other agents, the maximum number of iterations they can perform, and which tools they have access to.
Tasks represent specific pieces of work to be accomplished. Each task includes a description, the agent assigned to it, and the expected output format. Tasks can depend on the outputs of other tasks, creating workflows where information flows from one step to the next. CrewAI supports structured output through Pydantic models, ensuring that task results conform to a defined schema [4].
When a crew is executed, CrewAI orchestrates the agents according to the defined process type:
| Process Type | Behavior |
|---|---|
| Sequential | Tasks are executed in order. The output of each task is passed as context to the next task. This is the simplest and most predictable execution mode. |
| Hierarchical | A manager agent (automatically created or explicitly defined) delegates tasks to worker agents, reviews their outputs, and coordinates the overall workflow. This mimics a traditional management structure. |
| Parallel | Multiple tasks are executed simultaneously when they do not depend on each other, improving performance for workflows with independent sub-tasks. |
During execution, agents can communicate with each other, delegate work to other agents (if permitted), use their assigned tools, and iterate on their outputs. The framework includes built-in error handling and retry logic to manage failures gracefully [4].
CrewAI supports defining agents and tasks through YAML configuration files, separating the workflow logic from the application code. This approach makes it straightforward to modify agent roles, goals, and task descriptions without changing the underlying codebase. It also facilitates version control and collaboration on agent definitions [9].
A typical CrewAI project includes an agents.yaml file that specifies each agent's role, goal, backstory, and LLM, alongside a tasks.yaml file that defines each task's description, expected output, and assigned agent. This declarative configuration makes CrewAI one of the most beginner-friendly multi-agent frameworks available.
CrewAI provides a growing library of built-in tools and supports custom tool creation. The tool system is central to giving agents the ability to interact with external systems, retrieve information, and take actions in the real world.
The crewai-tools package ships with a wide range of pre-built tools:
| Tool Category | Examples |
|---|---|
| Web and search | Web search (via Serper, Google, Brave), web scraping, website reading |
| File operations | File reading, file writing, directory listing, PDF reading |
| Code | Code execution (Python), code interpreter, GitHub integration |
| Data | CSV search, JSON search, XML parsing, database queries |
| Communication | Email sending, Slack integration |
| RAG and knowledge | Document search, vector database queries, embeddings-based retrieval |
Developers can create custom tools using either a decorator pattern or a class-based approach. The decorator pattern allows wrapping any Python function as a tool with a simple @tool annotation, while the class-based approach extends the BaseTool class for more complex tools that need initialization parameters or state management [9].
Starting with the 1.0 GA release in October 2025, CrewAI introduced native support for the Model Context Protocol (MCP), an open protocol originally developed by Anthropic that standardizes how AI applications connect to external tools and data sources. MCP integration allows CrewAI agents to discover and use tools hosted on any MCP-compatible server, greatly expanding the ecosystem of available capabilities [10][11].
CrewAI supports MCP through two approaches:
| Approach | Description |
|---|---|
| Simple DSL (recommended) | The mcps field on agent definitions allows seamless tool integration using string URLs or structured configurations |
| MCPServerAdapter (advanced) | The crewai-tools library provides the MCPServerAdapter class for complex scenarios requiring manual connection management |
Three transport types are supported for MCP communication:
| Transport | Description |
|---|---|
| Stdio | Local server communication via standard input/output, suitable for tools running on the same machine |
| Server-Sent Events (SSE) | Unidirectional real-time streaming over HTTP for remote servers |
| Streamable HTTPS | Flexible bidirectional communication over HTTPS for production deployments |
MCP integration includes automatic tool discovery, name collision prevention through server prefixes, built-in timeout protection, and advanced tool filtering (both static and dynamic). For security, CrewAI documentation emphasizes validating MCP server origins and avoiding binding to all network interfaces to prevent DNS rebinding attacks [10].
In addition to MCP, CrewAI added native support for the Agent-to-Agent (A2A) protocol, enabling agents to delegate tasks, request information, and collaborate with remote agents built on different frameworks or platforms. A2A support also allows CrewAI agents to act as A2A-compliant server agents, meaning external systems can send tasks to CrewAI crews [12].
A2A communication supports multiple transport protocols including JSONRPC (the default), GRPC, and HTTP+JSON. Authentication schemes include Bearer tokens, OAuth2, API keys, and HTTP authentication. This capability is particularly valuable in enterprise environments where multiple agent systems from different teams or vendors need to interoperate [12].
The A2A integration is installed via pip install 'crewai[a2a]' and configured through A2AClientConfig (for connecting to remote agents) and A2AServerConfig (for exposing agents as servers).
CrewAI includes a sophisticated memory system that allows agents to retain and recall information across interactions. The memory architecture includes several types:
| Memory Type | Purpose |
|---|---|
| Short-term memory | Information relevant to the current crew execution, automatically managed |
| Long-term memory | Persistent information that survives across multiple crew executions |
| Entity memory | Structured information about specific entities (people, organizations, concepts) encountered during execution |
| User memory | Information specific to users interacting with the system |
The memory system uses an LLM to analyze content when saving, inferring scope, categories, and importance. Memories are organized into a hierarchical tree of scopes, similar to a filesystem, with each scope represented as a path (for example, /project/alpha or /agent/researcher/findings). Recall uses adaptive-depth scoring that blends semantic similarity, recency, and importance [13].
CrewAI Flows are a production-oriented orchestration layer introduced in 2025 that provides fine-grained control over complex automations. While crews enable autonomous collaboration between agents, Flows add structured, event-driven automation with precise control over execution paths [13][14].
Flows are designed for real-world production scenarios where reliability, error handling, and deterministic behavior are critical. They can combine single LLM calls, crew executions, and arbitrary Python code into coherent pipelines. Key capabilities of Flows include:
| Feature | Description |
|---|---|
| Event-driven execution | Actions triggered by events or external inputs such as API calls, file system changes, or webhooks |
| State management | Structured state that persists across flow steps, enabling complex multi-step workflows |
| Conditional branching | Dynamic decision-making based on intermediate results |
| Crew integration | Native support for embedding crew executions within flow steps |
| Error handling | Built-in retry logic, fallbacks, and exception management for production reliability |
| Event bus | Developers can create listeners that run when specific events are emitted; enterprise customers can register webhooks for these events |
As of early 2026, Flows process over 12 million executions per day across industries including finance, federal government, field operations, and manufacturing [14].
CrewAI provides built-in mechanisms for training and testing multi-agent systems, which help developers optimize agent performance before and during production deployment [15].
Training is initiated through the CLI with crewai train -n <n_iterations>. During training, the crew runs its workflow normally but requests human feedback at the end of each task. Learning is extracted from the feedback and the difference between initial and improved outputs, then stored in .pkl files. During normal execution, each agent automatically loads its consolidated suggestions and appends them as mandatory instructions to the task prompt, ensuring consistent improvements without modifying agent definitions [15].
Testing allows developers to run the system N times (via crewai test -n <n_iterations>), with each task result evaluated by an evaluator LLM that returns individual performance scores. This enables comparison across different LLMs to find the optimal balance of cost and quality.
The knowledge system allows agents to access and reason over external information sources, including documents, databases, and APIs. This enables retrieval-augmented generation (RAG) patterns where agents can ground their responses in specific data rather than relying solely on their underlying model's training data [9].
CrewAI supports guardrails that validate and constrain agent outputs at both the task and agent level. Guardrails can enforce formatting requirements, content policies, safety constraints, and business rules, helping ensure that agents produce outputs suitable for production environments.
Alongside the open-source framework, CrewAI offers an enterprise cloud platform launched in October 2024 with the Series A funding announcement. Initially called CrewAI Enterprise, the platform was later rebranded as CrewAI AMP (Agent Management Platform), providing a comprehensive infrastructure for building, deploying, and managing AI agents at scale [1][16].
The platform is organized around four main pillars:
| Pillar | Capabilities |
|---|---|
| Build | CrewAI Studio, a no-code visual editor plus AI copilot for designing agents; Python APIs for engineers; dozens of built-in integrations (Gmail, Slack, HubSpot, Salesforce, Notion, GitHub) and custom tool support |
| Observe | Real-time tracing of every agent action; automated and human-in-the-loop (HITL) training; security guardrails; LLM testing; hallucination score tracking |
| Manage | Centralized LLM management; role-based access control (RBAC); cron scheduling; usage dashboards; performance metrics; deployment history and streaming logs |
| Scale | Automatic serverless deployment on CrewAI cloud or private infrastructure (AWS, Azure, GCP); SOC2, SSO, and FedRAMP High compliance for enterprise customers |
CrewAI Enterprise integrates with all major LLM providers and hyperscalers. The v1.0 GA release added built-in tracing to every execution without requiring third-party observability tools, and a unified CLI for local development, staging tests, and production deployment [3].
CrewAI has partnered with Andrew Ng and DeepLearning.AI to produce a series of educational courses on multi-agent systems:
| Course | Description |
|---|---|
| Multi AI Agent Systems with crewAI | Introductory course teaching how to break down complex tasks into subtasks for multiple specialized agents |
| Practical Multi AI Agents and Advanced Use Cases with crewAI | Intermediate course covering real-world applications and advanced patterns |
| Design, Develop, and Deploy Multi-Agent Systems with CrewAI | Advanced course covering building agents with tools, memory, and guardrails; coordinating teams; and deploying systems with tracing and monitoring |
More than 230,000 people have completed the first two courses, making it one of the largest educational programs in the world for practical, production-ready AI agents [8][17].
CrewAI competes primarily with AutoGen (now part of the Microsoft Agent Framework), LangGraph (part of the LangChain ecosystem), and newer entrants such as OpenAI's Agents SDK. Each framework takes a fundamentally different approach to multi-agent orchestration.
| Dimension | CrewAI | LangGraph | AutoGen |
|---|---|---|---|
| Core paradigm | Role-based teams | Graph-based workflows | Conversation-driven agents |
| Configuration | YAML-driven, declarative | Code-first, programmatic | Code-first, programmatic |
| Agent definition | Role, goal, backstory (natural language) | Nodes and edges in a state graph | Conversable agents with system prompts |
| Execution model | Sequential, hierarchical, parallel | Directed graph with conditional edges | Message passing, group chat |
| State management | Crew-level shared context; Flow-level structured state | Persistent state with reducer logic for concurrent updates | Conversation history |
| MCP support | Native (since v1.0, October 2025) | Not natively supported (as of early 2026) | Not natively supported (as of early 2026) |
| A2A protocol support | Native (since 2025) | Not natively supported | Not natively supported |
| Learning curve | Low to moderate | Moderate to high | Moderate |
| Best for | Structured business workflows, clear role separation, rapid prototyping | Complex decision pipelines, branching logic, precise control flow | Conversational tasks, brainstorming, human-in-the-loop scenarios |
| Enterprise offering | CrewAI AMP (cloud platform with no-code studio) | LangSmith (monitoring and evaluation) | Microsoft Agent Framework (enterprise integration) |
| Production readiness | Battle-tested; 1.4B+ executions; used by Fortune 500 | Mature; persistent workflows supported; v1.0 reached late 2025 | Evolving; shifted to maintenance in favor of Microsoft Agent Framework |
| GitHub stars (March 2026) | ~45,900 | ~15,000+ | ~40,000+ |
CrewAI's primary advantage is its intuitive, role-based approach that allows non-technical stakeholders to understand and contribute to agent design. The YAML-driven configuration and natural-language agent definitions lower the barrier to entry significantly compared to LangGraph's graph theory concepts (nodes, edges, state schemas) or AutoGen's programmatic conversation patterns [18][19].
LangGraph offers greater flexibility for complex workflows with sophisticated branching, state management, and persistent checkpointing, but requires deeper technical investment. It reached v1.0 in late 2025 and has become the default runtime for all LangChain agents [18].
AutoGen excels at conversational, free-form interactions and has strong human-in-the-loop support. However, Microsoft shifted AutoGen to maintenance mode in favor of the broader Microsoft Agent Framework, which may affect its long-term trajectory [19].
A common pattern observed in the industry is to prototype in CrewAI for fast iteration, then rewrite the production path in LangGraph once the agent design is confirmed and precise control flow is needed [18].
CrewAI has been applied across a wide range of domains and industries:
| Use Case | Description |
|---|---|
| Content creation | Crews of researcher, writer, and editor agents collaborating to produce articles, reports, or marketing materials |
| Customer support | Triage agents routing inquiries to specialized agents for billing, technical support, or account management |
| Data analysis | Analyst agents processing data, generating visualizations, and producing insights reports |
| Software development | Agents handling code review, testing, documentation, and deployment tasks |
| Financial research | Research agents gathering market data, analyst agents interpreting trends, and reporting agents compiling findings with regulatory compliance |
| Recruitment and HR | Agents screening resumes, scheduling interviews, generating candidate summaries, and forecasting workforce needs |
| Sales and marketing | Agents integrating with CRM platforms to analyze customer sentiment, enrich marketing databases, predict consumer behavior, and generate targeted campaigns |
| Market research | Multi-agent systems that autonomously research, synthesize, and forecast market dynamics, providing real-time actionable insights |
CrewAI's GitHub repository has accumulated over 45,900 stars as of March 2026, with more than 5,900 forks and 250+ contributors, making it one of the most popular tools in the multi-agent space. The repository receives 1.8 million monthly downloads via PyPI [2][3].
The framework has been integrated with major cloud platforms and AI providers. Amazon Web Services (AWS) published official guidance for using CrewAI with Amazon Bedrock, and integrations exist for OpenAI, Anthropic, Google, and other model providers. The framework also supports local models through Ollama and other local inference solutions [20].
CrewAI's community maintains an active forum at community.crewai.com and contributes to a growing ecosystem of third-party tools, templates, and tutorials. Over 100,000 developers have completed certification through community courses at learn.crewai.com [3].
IBM, Microsoft, Procter & Gamble, Walmart, SAP, Adobe, and PayPal are among the major enterprises reported to use CrewAI for production agent workflows [3].
CrewAI hosts an annual conference called CrewAI Signal. The second edition, CrewAI Signal 2025, was held on November 20, 2025, in San Francisco and sold out with over 500 attendees from Fortune 500 teams. The event featured deep dives into real-world AI agent use cases, live demos, workshops, and networking with AI leaders. A satellite event was also held in Trivandrum, India, on November 17, 2025, reflecting CrewAI's global expansion [8].
Under the hood, CrewAI is built on Python and uses a modular architecture that separates agent definitions, task management, tool integration, and execution orchestration into distinct components. The framework supports both synchronous and asynchronous execution, with async support becoming increasingly important for production deployments that need to handle multiple concurrent crews [9].
CrewAI's orchestration engine manages workflows that adapt to task dependencies, supporting various execution models including sequential, parallel, and conditional processing. The engine handles agent communication through a shared context mechanism, where task outputs are automatically passed to downstream tasks as context [4].
The framework is model-agnostic, supporting any LLM through a unified interface. This allows developers to use different models for different agents within the same crew, optimizing for cost, speed, or capability as needed. For instance, a simple routing agent might use a smaller, faster model while a complex reasoning agent uses a more capable (and more expensive) one [9].
With the v1.0 GA release, CrewAI adopted a monorepo structure and locked its Crew and Flow APIs for long-term stability. Built-in tracing ships with every execution, providing instant observability without requiring external tools. The unified CLI consolidates local development, staging, and production deployment into a single interface [3].
As of March 2026, CrewAI is at version 1.10.1 and continues to be one of the leading multi-agent orchestration frameworks. The framework has achieved several notable milestones:
The multi-agent framework landscape has grown increasingly competitive, with new entrants including OpenAI's Agents SDK, Google's Agent Development Kit, and Amazon's Strands Agents joining established players like LangGraph and AutoGen (now Microsoft Agent Framework). CrewAI has maintained its position through its focus on simplicity, role-based design, broad protocol support, and enterprise readiness [18][19].
CrewAI's enterprise business continues to grow, supported by the $18 million in funding raised in 2024. The CrewAI AMP platform provides a complete lifecycle for building, observing, managing, and scaling AI agent systems, with compliance certifications including SOC2, SSO, and FedRAMP High [1][16].