What is the difference between an Agent and a Task in CrewAI?

An Agent is a persona — a role, a goal, and a backstory that define who is doing the work. A Task is a unit of work — a description and an expected output — assigned to an agent. Agents are reusable across tasks, and a crew is the collection of agents working through a list of tasks.

When should I use sequential versus hierarchical process?

Sequential runs tasks one after another, passing each output to the next, which is predictable and cheap. Hierarchical adds a manager LLM that decides which agent handles what and can re-delegate, which is more powerful for open-ended work but uses more tokens. Start sequential and upgrade only if you need dynamic delegation.

How do agents share information between tasks?

In a sequential process the output of one task is automatically provided as context to the next, so a researcher's findings flow into the writer's task without manual wiring. You can also set context explicitly on a task to pull in outputs from specific earlier tasks.

How do I give agents real tools like web search?

Pass a tools list to an Agent or Task. CrewAI ships integrations (search, scraping, file I/O) and supports custom tools — any Python callable wrapped as a tool. The generated script omits tools for clarity, but adding tools=[search_tool] to the researcher is the usual next step.

Does CrewAI only work with OpenAI?

No. CrewAI uses LiteLLM under the hood, so you can point it at Anthropic, Google, Groq, or a local model via Ollama by setting the model name and provider credentials. The OPENAI_MODEL_NAME and key in the script are just the most common starting point.

How to Use CrewAI to Build AI Agent Teams

What CrewAI is

CrewAI is a Python framework for orchestrating teams of AI agents that collaborate on a goal. Its mental model is deliberately human: you hire agents with roles, goals, and backstories, hand them tasks with clear expected outputs, and organise them into a crew that runs a process. That role-based framing makes multi-agent systems approachable — you design a team the way you’d staff a project. The builder above turns a few roles and a topic into a complete, runnable crew.

How it works

Three primitives do all the work:

Agent — a persona defined by role, goal, and backstory. These fields are not decoration; they are injected into the prompt and genuinely steer the agent’s behaviour and priorities.
Task — a description plus an expected_output, assigned to one agent. The expected output is what keeps results focused and chainable.
Crew — the agents and tasks bundled together with a process.

The process determines coordination. Process.sequential runs tasks in order, feeding each output forward as context — a researcher gathers facts, then a writer turns them into a report. Process.hierarchical introduces a manager LLM that dynamically assigns and reviews work, which is more flexible for ambiguous goals at the cost of more model calls.

Calling crew.kickoff() starts the run; CrewAI handles the message passing, context flow, and aggregation, returning the final task’s output.

Tips and a real example

The classic CrewAI starter is a research crew: a Senior Researcher whose goal is finding accurate, recent facts, and a Tech Writer who turns those facts into a polished report. Run sequentially, the researcher’s notes flow straight into the writer’s task. That two-role pattern (which the builder defaults to) is enough for blog drafts, market summaries, and briefings.

Write specific goals. “Find accurate, recent facts and sources” beats “research things” — the goal text is doing real prompting.
Make expected outputs concrete. Asking for “a polished report in markdown” produces far better structure than leaving it implicit.
Add tools when prose isn’t enough. Wire a web-search tool to the researcher so it grounds claims in live data rather than its training cutoff.
Stay sequential until you must scale up. Hierarchical crews are powerful but spend more tokens and are harder to debug; reach for them only when a fixed task order genuinely can’t express the work.