CrewAI

2.1 Scaffold your first CrewAI project

CrewAI is code, not a canvas. A crew is a team of agents — each with a role, a goal and a backstory — working through tasks. The CLI scaffolds the whole project for you: agents and tasks as YAML, the assembly as Python, so you start by reading a working example rather than an empty file.

Generate a runnable two‑agent CrewAI codebase with a single command

Run these in a terminal with Python 3.10+. crewai create crew scaffolds a whole project folder for you — you don’t write it from scratch. Prefer uv? uv tool install crewai installs the same CLI faster and keeps it out of your global Python.

1. Open a terminal and execute pip install crewai to add the framework to your environment
2. Run crewai create crew my_first_crew --classic to generate a project directory pre‑populated with a researcher and a writer agent using the classic YAML scaffold
3. Edit the generated .env file and insert your LLM API key so the agents can access a model

• You'll see A project folder appears containing researcher and writer agents together with configuration files
• Takeaway One CLI command creates a complete multi‑agent scaffold you can immediately inspect and run
• Check The scaffold produced a researcher and a writer. Where is each one actually defined?
• Cost The CrewAI framework is free and open-source (MIT). You only pay your LLM provider for the tokens the agents use when you run the crew — billed separately by them.

2.2 Inspect a crew’s configuration

A scaffolded crew consists of three files. agents.yaml defines each agent’s role, goal and backstory; tasks.yaml describes what each task should produce and which agent handles it; crew.py wires the YAML files together into a runnable Crew.

Do this first Scaffold your first CrewAI project

Identify how agents, tasks and the crew definition work together by editing the generated files

CrewAI separates who the agents are (agents.yaml), what they do (tasks.yaml), and how they're assembled (the Python crew.py). Edit the YAML — you rarely touch the Python at first.

1. Open config/agents.yaml to view each agent’s role, goal and backstory
2. Open config/tasks.yaml to see each task’s description, expected output and assigned agent
3. Open crew.py to observe how the YAML files are imported and assembled into a Crew

• You'll see The crew broken into separate YAML files for agents and tasks plus a short Python script that binds them
• Takeaway A crew is just agents + tasks + an assembly step; using YAML keeps the structure clear and easy to change
• Check You want an agent to stop writing in bullet points. Which of the three files do you edit?
• Cost Editing config files costs nothing — you only spend LLM tokens when you actually run the crew.

2.3 Execute a crew and observe agent reasoning

Running a crew prints its thinking, not just its answer. Each agent announces the task it took, reasons out loud, and hands its output to the next agent as context.

Do this first Inspect a crew’s configuration

Run your crew and see agents reason, hand off work, and generate a report

Run this from inside your project folder. Make sure your LLM API key is set in .env first, or the run will fail to reach a model.

1. Execute crewai run from the project folder
2. Monitor the handoff as the researcher’s output is passed to the writer agent
3. Read the final report printed in the terminal or written to a file

• You'll see Two agents printing their step‑by‑step reasoning, passing work between them, and producing a short written report from a single command
• Takeaway Running a crew makes the abstraction concrete: you see agents reason, collaborate and deliver
• Check The writer produced a report about the wrong topic. Which part of the run log tells you whether the researcher or the hand‑off was at fault?
• Cost This run spends real LLM tokens billed by your provider (a small report is typically a few cents). The CrewAI framework itself stays free.

2.4 Parameterise a crew: a market-data collector in YAML

In a scaffolded crew the agents and tasks are YAML files, not Python. Anything in {curly_braces} is an input the crew fills in at run time, turning a one‑off script into a reusable tool.

Do this first Execute a crew and observe agent reasoning

Define a collector agent in YAML with a {ticker} input and run the same crew against different companies

Run this in your terminal inside the project folder from Lesson 0. Swap TSLA for any ticker — the crew definition does not change.

1. Open config/agents.yaml and add a collector agent with a role, a goal containing {ticker} and a short backstory
2. Open config/tasks.yaml and add a collect_task that names agent: collector, describes the desired output and reuses the same {ticker} placeholder
3. Edit crew.py to include the new task in the crew’s tasks list so it runs
4. Run crewai run --input ticker=TSLA, then run it again with a different ticker and compare the two outputs

• You'll see The collector’s reasoning streams in the terminal, then a formatted summary of price, change and headlines for each ticker you passed
• Takeaway Put the changing part in {braces} and pass it as an input — that is the line between a script you rewrite and a crew you reuse
• Check Where does a crew get the value for {ticker} — the YAML, crew.py, or the command line?
• Cost CrewAI itself is free and open-source (MIT). Each run spends your own LLM tokens, and a two-agent crew makes several calls per run — so re-running it across ten tickers costs roughly ten times one run.

2.5 Add a second agent: an OCR extractor that reads a PDF

A crew grows by adding a pair of entries, one in agents.yaml and one in tasks.yaml. The framework then sequences them so the second agent receives the first’s output as context automatically.

Do this first Parameterise a crew: a market-data collector in YAML

Add an OCR extractor agent that turns a PDF into text and hands that text to the rest of the crew

Drop any PDF named sample.pdf into the project folder first, then run this in the terminal from that folder.

1. Add an extractor entry to config/agents.yaml with role "document extractor", a goal referencing {file} and a backstory stating it returns plain text only
2. In crew.py, attach CrewAI’s file‑read tool (e.g., FileReadTool) to the extractor agent so it can open PDFs
3. Add an extract_task entry to config/tasks.yaml pointing at agent: extractor and place it before any task that consumes the extracted text
4. Run crewai run --input file=sample.pdf and read the log to see the extractor invoked first and its output used as context for the next agent

• You'll see The log shows the extractor starting, calling its file tool, printing the extracted text, then the next agent’s turn opening with that text in its context
• Takeaway Extending a crew is two YAML entries plus a tool — the framework passes one agent’s output to the next
• Check Why does an extractor agent without a file‑reading tool still produce an answer, and why is that answer worthless?
• Cost Free framework; the cost is tokens. OCR-style tasks push a whole document into the prompt, so a long PDF is by far the most expensive step in the crew — check the page count before you run it repeatedly.

2.6 Equip agents with custom tools and select their LLM

A tool is a callable function agents can use; without one they only talk. The LLM setting is per‑agent, allowing different models for gathering versus writing.

Do this first Add a second agent: an OCR extractor that reads a PDF

Make agents more capable by attaching tools and configuring the language model they use

CrewAI is model-agnostic and ships a large library of tools/integrations. You attach tools to agents in code; you pick the model via your .env / config.

1. Install the extra tools package with pip install 'crewai[tools]' if you need built‑in utilities
2. Add a tool (e.g., a search or file‑reading utility from CrewAI’s toolkit) to the desired agent definition in config/agents.yaml
3. Specify the LLM for each agent by adding an llm field with your provider name and API key in the agent’s YAML entry
4. Run the crew with crewai run and observe the researcher call its tool mid‑task and pass the result to the writer

• You'll see An agent invokes a real tool during its reasoning and the crew runs on the chosen LLM
• Takeaway Tools turn agents from talkers into doers and model‑agnostic configuration avoids vendor lock‑in
• Check Why is it usually wrong to give every agent in a crew the same model?
• Cost Still free at the framework level; tools that hit paid APIs and your LLM usage are billed by those providers. Bringing your own key means you control the spend.

2.7 Pick the right structure for your AI project

CrewAI offers two structures: a Crew where agents self‑organise, suited for open‑ended research; and a Flow with explicit steps and branching, suited for deterministic pipelines.

Do this first Equip agents with custom tools and select their LLM

Decide whether to use an autonomous crew or a deterministic flow for a given task

CrewAI gives you two paradigms. A Crew is autonomous — agents decide how to collaborate. A Flow is structured and event-driven — you control the exact path.

1. Identify whether the task requires open‑ended research or a fixed sequence of actions
2. Select a Crew if agents should self‑organise their collaboration
3. Select a Flow if you need deterministic, event‑driven control with explicit steps and branching

• You'll see A clear mental model for choosing between an autonomous crew and a structured flow for a given problem
• Takeaway Reach for Crews when you want emergent collaboration, and Flows when you need a controlled, repeatable pipeline
• Check A nightly report that must always contain the same five sections — Crew or Flow?
• Cost Conceptual lesson — no extra cost beyond whatever you run while experimenting.

2.8 Deploy your crew from GitHub in CrewAI AMP

AMP is CrewAI’s hosted platform; it pulls your crew from a GitHub repository and runs it on managed infrastructure, keeping environment variables separate from source code.

Do this first Pick the right structure for your AI project

Push a code‑first crew to a repository and launch it on the hosted platform with one click

In the CrewAI AMP UI, select Deploy your crews from GitHub, enter the repo URL and choose the branch as shown, then add any required environment variables and optionally enable Automatically deploy new commits. Watch for the deployment status to turn green indicating success.

12 Enlarge

1. 1 Which crew ships empty until you connect GitHub
2. 2 Which branch ships one branch, not the whole repo

Best viewed on desktop — tap Enlarge to read the numbered controls.

1. Push your crew to a GitHub repository
2. In CrewAI AMP, choose Deploy your crews from GitHub, select the repository and branch, then click Deploy
3. Enter required environment variables (e.g., your LLM API key) and optionally enable automatic redeploy on new commits

• You'll see Your crew appears in CrewAI AMP as a deployed workflow linked to the selected GitHub repo
• Takeaway A visual editor and GitHub integration let you turn code into a hosted agent without managing servers
• Check Why does deploying from GitHub, rather than uploading files, matter for a crew you will keep changing?
• Cost AMP's Basic tier is free (50 workflow executions/month, visual editor, AI copilot, GitHub integration). Enterprise is custom-quote (50 hours of development/month, hosted or private infra). You still bring your own LLM key — that usage is billed separately.

2.9 Track and manage deployed crews

A deployed crew is exposed as an HTTP endpoint with a bearer token. The dashboard lists each crew’s URL and token and provides controls to rotate tokens, redeploy or delete the crew.

Do this first Deploy your crew from GitHub in CrewAI AMP

Identify each crew’s URL and bearer token and control them from a single dashboard

Run this in a terminal with Python 3.10+. When the progress view finishes, note the live URL and Bearer Token shown, then open the management dashboard (e.g., via crewai dashboard) to verify your crew appears with its credentials.

1. Open the management dashboard to view all deployed crews side by side
2. Click Reset next to a token to rotate it
3. Use the Manage / Re‑deploy / Delete controls to modify or remove a crew

• You'll see A production dashboard listing your deployed crews with their URLs, bearer tokens and management controls
• Takeaway Deployments provide a secure endpoint that can be monitored and managed centrally
• Check Your crew’s URL is in a public repo. What is the one thing to do first?
• Cost Included in the AMP tier you're on (Basic counts runs against its 50/month). Calling your crew still spends your own LLM tokens, billed by your provider.