The open source Crewai framework allows you to quickly create teams of specialized agents to perform complex tasks with AI.
The agentic AI is at our door, harassment the publishers of models all day long, transforming this promise into an omnipresent refrain of tech conferences and marketing campaigns. However, it is still quite complex in 2025 to develop a real agental strategy without depending on a SaaS service or a publisher. Faced with this challenge, Crewai emerges as a particularly relevant solution. Designed to accelerate the adoption of AI agents, the tool responds precisely to the needs of companies that wish to develop their own agentic pipeline internally without redeveloping the entire underlying technological stack.
An orchestration on the company model
Launched in open source by João Moura, former Director of AI at Clearbit / Hubspot, Crewai is based on a principle similar to that of a traditional company: a clear hierarchy with a coordinator and specialized workers. The fundamental difference lies in the fact that workers are autonomous AI agents, each with specific skills and tools. A structure that seems to be inspired by the mode Scrum (Agile methodology where a self-organized team works by short iterations with a scrum master facilitator), but applied to IA agents. The central “crew” plays the role of conductor who pilots all agents, defines the global objectives and ensures the consistency of interactions between the various stakeholders.
Concretely, when a developer implements Crewai, he begins by defining specialized agents – for example a researcher who extracts web data, an analyst who processes this information and an editor who produces final content. Each agent is configured with a specific profile (level of expertise, personality, objectives) and is awarded tools such as access to external APIs, research capacities or data processing functions. The developer must then configure a workflow to structure collaboration: the researcher can transmit his results to the analyst who, after treatment, communicates them to the editor to produce the final text. An orchestration which, theoretically, makes it possible to automate complex tasks which would normally require several human workers.
How to create your crew?
Crewai settles quickly under Python with the PIP manager. (Pip Install ‘Crewai[tools]’). Crewai is compatible with OPENAI (GPT-3.5, GPT-4), Anthropic (Claude), Cohere, Mistral AI, as well as Ollama and LM Studio for the execution of local models. The configuration of the models is fairly simple: in installation just enter the API key of your model supplier, Crewai takes care of the rest.
Once the bookstore is properly installed, it is quite easy to create your first crew. The command to create your project is quite simple: Crewai Create Crew Nomduprojet. The framework then built a basic project with all components. Crewai revolves around four essential modules: agents, tasks, tools and crew, which orchestrates the entire system.
Agents
These are the main players in the system, built from the API of AI models. Each agent has a defined role (researcher, analyst, editor, etc.), a specific objective and tools to accomplish his task. The definition of an agent is done with a few lines of code, in YAML (a simplified XML language).
The tasks
The tasks represent the missions assigned to the agents. Each task has a description, an assigned agent and execution parameters.
Tools
The tools allow agents to interact with the outside world. Crewai natively supports integration with APIs like snow snow for web or tavily research for information for example. By default, Crewai supports dozens of different tools. It is also possible to add your own tools by hand.
The crew
It is the entity that orchestrates the entire system, bringing together agents and tasks in a structured process. The process can be configured according to several modes: sequential (the tasks run in a predefined order), hierarchical (an agent manager delegates to others) or consensual (agents collaborate to reach a consensus on decisions). Technically, communication is carried out by sharing a history of messages between agents. Each agent can access the results of the previous tasks and the reasoning that led to these results. For example, in a hierarchical mode, the agent manager can examine the work of his subordinates, ask for clarifications or improvements via direct messages, then integrate the responses in his final decision -making.
In the case above, when a user launches the CREW, the researcher performs his search on the requested subject and produces a list of 10 key points. This information is then transmitted to the report analyst who transforms them into a document structured in Markdown format. Everything runs sequentially and automated.
Simple or complex use cases
Crewai finds its place in many cases of professional uses where automation saves time. For example, for competitive intelligence, a team of agents can be configured to monitor competitors on different channels daily, analyze changes in strategy or price, and produce a synthetic weekly report. As part of a customer service, CREWAI can make it possible to set up an automated response system where an agent analyzes incoming demand, search for the appropriate solutions in the knowledge base, and generates a personalized response that the support team can quickly validate and send. Another use case: post-reunion management. From a meeting transcription, an agent can identify action points, another can structure them by priority and responsible, while a third is responsible for automatically send personalized monitoring emails to each participant with their assigned tasks and deadlines.
If these applications are attractive on paper, their implementation with Crewai still requires significant technical skills and several hours of development – to assess whether the game is worth the candle, the site cvlc.fr also offers a practical calculator which considers the time saved by the automation of a task. Despite the simplified approach in Yaml and Python, we are still far from a real low-code solution accessible to non-developmenters. Finally, the fine understanding of prompts and their final results is also far from being acquired in all development teams … until the day when the agent will be able to code his own colleagues.