Using Agent mode

Agent mode lets you delegate multi-step coding tasks to AI. Describe what you need in plain language — Agent reads your codebase, plans the work, edits files, runs tests, and shows you every change for review.

Opening Agent

Press Cmd + Shift + A or click the Agent icon in the sidebar. Type your task in the prompt field and press Enter to start.

Writing effective prompts

Be specific about what you want. Good prompts include the scope, expected behavior, and any constraints.

• Good: "Add error handling to all routes in src/api/ using our existing ApiError class"
• Vague: "Fix the API"

What Agent does step by step

1. 1

   Explores — Reads relevant files and understands your project patterns.

2. 2

   Plans — Creates a checklist of files to read, edit, and test.

3. 3

   Executes — Makes edits and runs terminal commands (with your approval).

4. 4

   Reports — Shows a full diff for you to accept or reject.

Agent never applies changes without your review. You can accept all, reject all, or pick individual file diffs.

Approving terminal commands

When Agent needs to run a command (e.g. npm test), it shows the command and waits for approval. Click Run to allow or Deny to skip. You can enable auto-approve for trusted commands in Settings.

Learn more

For a full feature overview with before/after examples, see the Agent mode product page.