Build your own agents

What you'll see in this lesson

1. An agent from the inside: a loop with tools

   An agent is a model that, in a loop, decides which tool to use, gets the result and keeps going until the task is done.

2. The minimal loop, by hand, with the Claude API

   Ten lines are enough: while the model asks for tools, you run them and send back the results.

3. Claude Agent SDK: the engine of Claude Code

   Anthropic's SDK gives you Claude Code's loop, tools, permissions and hooks as a library.

4. OpenAI Agents SDK: lightweight, with handoffs

   You create an Agent with instructions and tools, and the Runner runs the loop; a decorator turns a function into a tool.

5. Google ADK: from one agent to a system of agents

   Google's Agent Development Kit is optimized for Gemini, supports other models and includes a web interface for testing.

6. LangGraph: stateful workflows you can pause

   You draw the process as a graph of steps that share state, with branches, retries and pauses for a person to approve.

7. CrewAI: teams of agents with roles

   You define agents as if you were hiring people, with a role, a goal and a backstory, and CrewAI coordinates who does each task.

8. Which one to use in each case

   Start with the simplest thing that works: your main model's SDK, and a framework only when the workflow calls for it.

9. Summary: build your own agents

   Deep down, every SDK does the same thing: a loop of model and tools; what changes is the parts they give you ready-made.