Deep Agents Workshop

Why it matters

A plan that lives only in the model’s reasoning disappears under tool output. A deep agent keeps its plan as a document it updates between steps, which is what makes long tasks inspectable and repairable.

What to practise

Build the agent in deep_agent.py with create_deep_agent(): configure the model, the tools, the system prompt and the backend — FilesystemBackend, LocalShellBackend or a Docker sandbox. In the workshop’s UI, give it a multi-step request and watch write_todos record tasks as pending, in_progress or completed, and the task tool hand a subtask to a subagent. Then rebuild with Sandbox Mode on, ask what is in the workspace again, and note that the Docker container has no host mounts or network while web search and RAG still run in the application.

How it fits

The to-do list is the plan representation 5.3.2 asks you to attach checks to: a task marked completed is an action taken, not yet an outcome verified. Delegation to subagents is 5.3.1’s hierarchical planning.