Learn how Claude agents work
This course is for complete beginners. It explains what an AI agent is, how it uses tools and skills, and five common ways to organise agent work. You don't need any programming experience to follow it.
An AI model such as Claude reads text and writes text. An agent is a program that lets the model work towards a goal on its own. It can choose actions, use tools (for example "search the web" or "read a file"), look at the results and decide what to do next, repeating until the job is done.
Tool use and Agent Skills are real Claude features, documented by Anthropic. Sequential, Parallel, Route, Human in the Loop and Orchestrate are general design patterns: you build them with ordinary code around Claude. They are not buttons or switches in the Claude API. Anthropic describes several of them in its article Building effective agents. Each lesson shows which kind it is.
Your progress
Loading progress…
Recommended learning path
This is a suggested order. Each lesson explains the terms it uses, so you can start anywhere. Tool and Skill are the building blocks, the next four are ways to arrange work, and Orchestrate brings everything together. To track progress, press Mark as completed at the bottom of a lesson. Progress is saved in this browser only.
Topics
How agent workflows fit together
Think of the seven topics as layers:
🧱 Building blocks
Tools let the agent do things, such as search or send. Skills give the agent know-how: packaged instructions for a type of task.
🧭 Arranging the work
Sequential: do steps in order. Parallel: do independent steps at the same time. Route: pick the right path for each request.
🛡️ Staying safe
Human in the Loop lets a person approve, change or stop actions that are risky or can't be undone.
🎼 Putting it together
Orchestrate: a coordinator plans a large goal, hands the parts to specialists and checks the results.
Comparison table
| Concept | In one line | Use it when… | Avoid when… | What kind of thing it is |
|---|---|---|---|---|
| Tool | A function the agent can ask your code to run | The answer needs live data or a real action | Claude can already answer from the text you gave it | Claude API feature (tool use) |
| Skill | A reusable folder of instructions and resources | You repeat the same specialised task often | It's a one-off request, so a normal prompt is enough | Claude feature (Agent Skills) |
| Sequential | Steps run one after another | Each step needs the previous step's output | Steps don't depend on each other (use Parallel) | General pattern ("prompt chaining") |
| Parallel | Independent tasks run at the same time | Tasks are separate and speed matters | One task needs another's result | General pattern ("parallelization") |
| Route | Classify, then send to the right handler | Requests come in clearly different types | Every request is handled the same way | General pattern ("routing") |
| Human in the Loop | A person approves before key actions | Actions cost money, can't be undone or are public | The action is harmless and easy to reverse | General pattern, supported by permission features in some Claude products |
| Orchestrate | A coordinator plans and delegates to specialists | A big goal needs different kinds of expertise | One agent can do the job, which is simpler and cheaper | General pattern ("orchestrator-workers") |
Beginner project: Smart Support Inbox
Goal: design, on paper or in code, an assistant that helps a small online shop answer customer emails. The project uses every concept in the course.
- Route: sort each email into order question, refund request, product question or unclear. Unclear emails get a polite reply asking for more detail.
- Tool: define a
get_order_status(order_id)tool that looks up a pretend order list. - Skill: write a
SKILL.mdcalled "support-reply-style" that describes the shop's tone, sign-off and refund policy. - Parallel: when an email asks several separate questions, look up each one at the same time.
- Sequential: draft the reply, then check it against the style skill, then make the final version.
- Human in the Loop: every refund over $50 and every outgoing email waits for a person to click Approve.
- Orchestrate: a coordinator runs steps 1–6 for a batch of emails and produces a summary for the shop owner.
You're done when: you can trace three sample emails through your design and explain which pattern handles each step and why. Each lesson has a smaller mini project that builds one of these pieces.
Official documentation
- Claude documentation: the home for Claude product and developer docs
- Claude API (Developer Platform) docs: the Messages API, tool use, skills and SDKs
- Anthropic documentation: the original docs address, which points to the current docs
- Building effective agents: Anthropic's article on agent workflow patterns