Documentation / The agent

Ask for a task

How to work with the agent from start to finish without losing control.

Write what you need the way you'd tell a colleague. The agent proposes, the IDE controls, and you approve.

Before you start

  • Open the project and wait for indexing to finish (the editor's top bar shows it). You can type in the meantime.
  • Connect an AI: your subscription or an API key. See Connect an AI.
  • Show the chat with Ctrl+Alt+I if it's hidden. From the editor, Ctrl+L moves focus to the chat.

1. Describe the outcome

Say what you want and how you'll know it's right.

Instead of… Write…
“Fix the login” “Login fails for users with an accent in their email. Fix it and add a test that covers it.”
“Make an endpoint” “Add GET /api/orders/{id} that returns 404 if it doesn't exist. Follow the style of ProductController.”

You can attach files with the paper clip and mention code. Pasted code is recognized along with its language.

The agent always replies in the language you write to it in.

2. Answer the questions

If something is missing, the agent asks during analysis, all at once, in Questions from the AI cards:

  • Each question comes with its Recommended option.
  • Another answer… lets you write your own.
  • Let the AI decide delegates that question; You decide everything picks the recommended option for all of them.
  • Keyboard: 1–9 selects, Enter continues, ← goes back.

3. Follow the plan

Multi-file tasks become a mission with a plan of verifiable steps and live progress. The mission bar, above the chat, shows the current phase: Phase n of 8. See Missions and phases.

While it works, the workboard shows what it's doing at every moment: Analyzing, Investigating, Editing, Verifying…

You can keep typing: your messages are queued, and you can ask “how's it going?”.

4. Approve what has side effects

When the agent needs to run a command or do something with consequences, the Approval required card appears:

  • Approve or Run approves it; Reject discards it.
  • Always in this chat stops it from asking again about that same action in the conversation.
  • Ctrl+Alt+Enter approves everything pending.

5. Review the changes

No change reaches the disk without review. Each proposal arrives as a card with its diff. Open them all in AI changes with Ctrl+Alt+R. See Review changes.

6. Receive the delivery

When it's done, the agent:

  1. Measures the code it touched against the previous version and gives a single-line verdict.
  2. Builds, tests, and starts the application.
  3. Summarizes what was done, what was verified, the decisions, and what's pending.

Undo

  • Alt+Shift+Z undoes the last thing the agent applied.
  • Ctrl+Alt+Shift+Z undoes the last complete task, with all its files.
  • ↺ Undo task in the turn summary does the same, with a preview.

What the agent never does unless you ask

  • Commit, push, merge, or create branches.
  • Write to a database.
  • Read .env files, keys, or credentials.

See Sandbox and permissions.

More ways to work

If you want… Use
It to work on its own until it's done Unattended mode
To split a large job among several agents Agent team
To delegate something long without touching your working copy Background tasks (Alt+Shift+B)
To design a screen Web Studio