Lesson 6. Agent mode — AI works independently, you control

Topic breakdown

Agent mode (Cmd+I or Composer) is Cursor's most powerful mode. In standard Chat, the AI gives suggestions and you click Apply. In Agent mode, the AI acts independently: it opens files, creates new ones, makes changes, and even executes commands like 'npm install' in the terminal.

To launch Agent mode, press Cmd+I (Mac) or Ctrl+I (Windows) and describe your task. For example: 'Create a React project with a login page, dashboard, and settings. Add routing and a base layout for each page.' The AI will do this automatically.

Important: in Agent mode, the AI shows every step — which file it creates, what it modifies, what command it runs. At every step, you can click 'Accept' or 'Reject'. It is not completely autonomous — you remain in control.

Agent mode works best: when starting a new project, when you need to create multiple files, during code refactoring, and when fixing bugs. For simple one-line changes, Tab or Inline Edit is enough.

What you'll learn

• Launching and managing Agent mode
• Properly assigning large tasks to Agent (effective prompting)
• Controlling every step of Agent (Accept/Reject)
• Course-correcting when Agent makes a mistake
• Distinguishing tasks suitable and unsuitable for Agent
• Authorizing terminal commands and file creation

Deep dive

Technical workflow of Agent mode: when you describe a task, Cursor sends the full project context (file tree, contents of key files) and your request to the AI model. The model makes a plan (which files to create/edit, which commands to run) and executes it step-by-step. Between steps, the model sees the result of its previous actions.

Practical difference between Agent mode and Chat mode: in Chat you write 5 separate messages and click Apply 5 times. In Agent, you write 1 detailed task, and the AI executes all 5 steps itself. The result is the same, but Agent saves 5 times the effort. However, if Agent makes a mistake, it's harder to fix because it alters multiple files simultaneously.

Best practices for Agent mode: 1) Describe project structure before the task. 2) Specify the tech stack clearly. 3) Always test the result (open in browser, run in terminal). 4) If it fails 3+ times — split the task into pieces. 5) Write project rules in the .cursorrules file — Agent will obey them without question.

Agent mode is especially useful in vibe coding: you can build a startup MVP in an evening. For instance, a landing page + Telegram bot + admin panel — with Agent mode, this is 2-3 hours of work. Vibe coding democratizes development — making technology accessible to everyone.

Why it works

Common mistakes

• Giving Agent too vague a task — not 'make a website', but exact structure and functionality
• Blindly clicking Accept — review every file, click Reject on mistakes
• Starting over completely when Agent fails — it's better to explain the error and continue
• Asking for too large a project at once — do it in stages: structure first, then pages, then styling
• Not checking terminal commands — Agent rarely, but sometimes, might suggest dangerous commands

Lesson FAQ

Next step