Working with Agent

Press Cmd+I (Mac) or Ctrl+I (Windows/Linux) to open Agent in the side panel. Describe a task in plain words: "the signup form accepts an empty email; add validation and a test". Agent then works on its own in steps, using tools:

• search files and folders, and search the web,
• read files (including images, for models that understand them),
• edit files,
• run shell commands in a terminal and read their output,
• control a built-in browser to check a page, and ask you clarifying questions.

It keeps going, step after step, until it judges the task done. Each step is shown in the chat, so you can follow along.

You stay in charge while it works:

• Press Enter to queue a message; queued messages run in order after the current task.
• Press Cmd/Ctrl+Enter to send immediately, or use Send now, to steer the running agent; it reads the message at its next tool call.
• Stop it if it heads the wrong way, then say what to change.

Before Agent runs a terminal command, your run mode decides whether it asks you, checks the command automatically, or just runs it. A later lesson covers these settings; until then, read commands before approving them.

Good first tasks are small and checkable: explain the codebase, suggest three low-risk improvements and pick one, fix a failing test. Large, vague requests ("refactor everything") give the agent too much freedom and produce diffs that are hard to review. Before bigger changes, use Plan Mode (covered next level).