CrewAI Examples: Official Projects, Cookbooks and How to Run Them

CrewAI's official examples sit in four places, and together they cover more than the scattered pages suggest. The docs examples page lists four crews (marketing strategy, surprise trip, recruitment, job posting), three flows (lead score with a human in the loop, email auto responder, meeting assistant) and an example that connects CrewAI to LangGraph ^[1]. The CrewAI-examples repository adds a landing page generator, a trip planner, stock analysis and job description writing. Six cookbook notebooks each teach one technique. Start smaller, though. The quickstart and the two first-project guides are the examples to run first: the first-crew guide uses today's JSON-first crew layout, while the quickstart and the first-flow guide are Flow projects scaffolded with crewai create flow.

Where the official examples live and what each one teaches

The split between crews and flows is the useful part of that list. CrewAI's docs recommend a crew when the path to the answer is open-ended and a flow when you control the steps, their order and the conditions. The marketing strategy and surprise trip crews show agents dividing open research and writing. Lead Score shows a router sending work down different paths after a person gives feedback. Write a Book is the one to study when one crew's output must feed the next, because it chains crews inside a single flow ^[2].

The cookbooks are narrower. Each notebook isolates one technique, such as agents collaborating, crew planning, reasoning, or a task guardrail that checks output before it is accepted, and two are tied to Gemini features: search grounding and video summaries ^[3]. They suit a team that already has a crew running and wants to add one capability without reading a whole project. Read the reasoning topic against current settings: the agent-level reasoning flag is deprecated in favor of a planning configuration, so treat code that still sets it as dated.

On the paid side, CrewAI AMP (Agent Management Platform) offers agentic workflow templates on both the Basic and the Enterprise plan, and its Visual Agent Builder includes a template library for building agents without code. The free Basic plan covers 2 automations and 50 workflow executions a month on CrewAI cloud, with no additional executions. By our arithmetic that is about 1.7 a day on a 30-day month. Enough to try a template. Too little to run one all day.

The three starter projects to run before anything else

The quickstart is the shortest working example. It builds a Flow whose state holds a topic and a report, then runs a one-agent crew that searches the web and writes a markdown report to disk. You create it with crewai create flow, install it with crewai install and start it with crewai run. It needs two keys: one for your LLM provider and a Serper.dev key, set as SERPER_API_KEY, for web search. Both providers bill you directly. CrewAI does not.

The first-crew guide is the crew-only version. A researcher agent uses SerperDevTool, a later task receives its output through context, and the final result lands in output/report.md. The project is JSON-first: agents live in agents/*.jsonc files, tasks and crew settings in crew.jsonc, and crewai run loads them directly. That layout is now the default. The older crew.py plus YAML layout still works behind the --classic flag, so a classic project you find elsewhere is older rather than broken.

The first-flow guide teaches the most. It takes user input, makes a direct LLM call that returns a structured Pydantic response, then hands the result to a writer and a reviewer agent running in sequence, all joined by @start and @listen and a typed flow state. It expects basic Python. One catch matters here. Its outline call uses an OpenAI model, so it needs an OpenAI key even when your agents run on another provider. Every crewai create flow project also ships a working poem_crew to take apart.

Getting an older example to run on today's CrewAI

The environment comes first. CrewAI needs Python 3.10 or newer and below 3.14, uses uv for dependencies, and installs its command line with uv tool install crewai. Inside any example, run crewai install before crewai run, from the project root. Missing {placeholder} inputs are asked for at run time, and crewai run checks each model's API key and prompts for missing ones, saving them to .env. Flows take JSON inputs through --inputs, and crewai flow plot draws the flow as an HTML page you can read before touching the code.

Three changes catch older examples. CodeInterpreterTool was removed in April 2026, so an example that ran code with it needs an E2B or Daytona sandbox tool instead, each with its own account. Tool-result caching became opt-in in July 2026, so do not assume a repeated tool call is served from cache. Memory was rebuilt in March 2026 into one unified Memory class, which replaces the separate short-term, long-term and entity memories that older code configures. Upgrades need care too, because crewai install only syncs the lock file; uv add is what moves versions forward.

Turning an example into your own crew or flow

Adapting starts with the agents. Each one is a role, a goal, a backstory and a set of tools, so rename them for your job, swap the research tools for the ones your data needs, and keep the task descriptions and expected outputs concrete. Keep the process sequential at first. The docs advise moving to a hierarchical process, with a manager model or manager agent, only when the workflow's complexity calls for it. Delegation between agents is off by default, and that is the safer setting for specialists.

Structure comes next. The docs recommend starting a production app with a Flow and running a crew inside one of its steps, and they call a direct LLM call more efficient than a crew when you need a single structured output. Add checks before trusting results: task guardrails can be Python functions or plain-language rules checked by the agent's LLM, and a failed check retries the task up to three times by default. For human review, set human_input on a task, or use @human_feedback in a flow (CrewAI 1.8.0 or later) to pause and route on what the reviewer says. Our CrewAI Flows guide covers routers, state and persistence, and the CrewAI tutorial walks through setup from zero.

When the adapted project is ready, crewai deploy create sends it to AMP from a git origin or as a ZIP, and the first deploy takes about a minute. AMP needs a committed uv.lock file, so commit it before deploying. The same command family covers status, logs, push, list and remove, and AMP can also deploy straight from GitHub, any Git provider or Studio. Once deployed, the example becomes an automation with its own API, triggers and traces.

Architectures worth copying from CrewAI's case studies

These are vendor accounts, not downloadable code. Their value is the agent roles. Each one shows how a working team divided a job, which is the hardest part of adapting any example. Read the outcomes as what the customers report, not as benchmarks, and copy the structure: which agent researches, which one writes, which one checks, and where a person still signs off before anything leaves the system.

Sales outreach and lead scoring

CrewAI's case study says DocuSign runs outreach as a deterministic Flow that controls sequencing, errors and state, with five agents inside it: an Identifier, a Researcher, a Composer, a Validator and an Orchestrator. The agents pull data from Salesforce and Snowflake and apply business rules to judge fit. The Validator checks quality, hallucination risk and personalization before a message goes out or reaches a person for review. Brickell Digital, an AI consultancy, built a smaller version: an agent that scrapes fundraising databases and VC networks and scores prospects against an ideal customer profile. Our CrewAI for sales teams article maps that pattern onto a CRM.

Documents, data and operations

CrewAI's case study says Gelato embeds CrewAI agents in Gelato Connect to map product SKUs and validate uploads, and that Gelato reports onboarding a new logistics carrier in 10 minutes instead of 5 days ^[4]. IBM Consulting, by CrewAI's account, extracts applicant data from documents, summarizes the findings and calculates program eligibility, combining Flows with crew agents. A healthcare staffing company, in a CrewAI blog account, automated physician credentialing checks (license, medical board and sanction lists) with an agent that pushes JSON results to Snowflake. All three share one shape: read a document, check it against rules, write a structured record.

Content, travel and support

General Assembly, CrewAI says, used a crew to generate lesson content and instructor guides. Waynabox, a surprise-trip travel company, combines web search, weather, preference and communication agents to build one itinerary per traveler and send it over WhatsApp. An HR services company started with a person reviewing every answer from its helpdesk agent and, by CrewAI's account, now lets the agent handle part of the tickets without review. That last one is the pattern to copy. Start supervised, then loosen.

Which CrewAI example to open first

Open the quickstart first, because it touches a Flow, a crew, a tool and a saved report in one small project, and it needs only two keys: one for your LLM provider and one for Serper.dev. Then pick by job shape. A job with an open-ended research or writing core points to the surprise trip or marketing strategy crews. A job with fixed steps, branches and an approval points to Lead Score or Email Auto Responder. A job that reads documents and checks them against rules points to the IBM and credentialing architectures, rebuilt as a flow with one crew inside.