How Do You Redirect Cursor's Agent While It's Working?
You don't have to wait for Agent to finish. Press Enter to queue your next instruction, press Cmd+Enter (Ctrl+Enter) to send it immediately, or click Stop or press Cmd+Shift+Backspace to cancel the run. Choose based on whether the current work is still useful.
Correction for your current work: [the one fact or constraint it's missing]. Keep what you've done so far unless it conflicts with this. Tell me in one sentence what changes as a result, then continue. For a long-running objective, in a new chat: /goal [objective, for example: fix all flaky tests and make CI green]
- Queue (Enter)
- Send now (Cmd+Enter)
- Stop
Try it now
- 1Give Agent a task that takes a few minutes, such as adding tests for one module.
- 2While it works, type a second small request and press Enter. It appears in order below the active task.
- 3Type a short correction and press Cmd+Enter (Ctrl+Enter). It worked if the agent picks up the correction before the current task ends, and the queued message runs afterward.
Use it when
Queue when the next task doesn't depend on seeing this result in detail. Steer when you notice something it needs to know, such as "use the v2 API". Stop when you can see it's going wrong and every further edit will only need undoing.
Skip it when
Don't steer around a fundamentally wrong plan. Patching with messages keeps the bad work in place, and Cursor's advice in that case is to revert and refine the plan. Don't pile on a long queue either, because later messages were written without knowing how the earlier ones turned out.
Why it works
These act at different points in the agent's loop. A queued message runs after the current task finishes, and queued items run in order. Cmd+Enter appends your message to your most recent one and attaches it to tool results, so it's processed right away. Newer steering works similarly: type a follow-up and click Send now, or press Enter twice, and it arrives at the agent's next tool call, so it doesn't cut off work mid-action. Cursor's docs say steering is on cursor.com/agents now and rolling out in the Agents Window.
Example
Say Agent is adding tests to auth.ts and you see it importing Jest when your repo uses Vitest. You press Cmd+Enter and send "Tests use Vitest, see src/__tests__/example.test.ts". It adjusts and carries on. You also queue "then add the same coverage to session.ts" with Enter. Afterward you add the Vitest fact to a rule, so you don't have to say it again.
Going further
Agent reads each message as a new job, unless you use /goal, which gives it an objective to work toward until it's complete. Pair it with a Custom Mode to follow a playbook, or use /loop to repeat a prompt at intervals. Cursor's docs say /goal is rolling out, so try a new chat if you don't see it. In the CLI the keys differ: Enter steers at a safe boundary, a second Enter interrupts the turn, and Ctrl+C pauses a goal. And you can always open a side chat to ask a question without interrupting anything.
Common mistakes
- Pressing Enter on an urgent correction. It queues by default, so the agent won't see it until the task is over.
- Pressing Cmd+Shift+Backspace twice. The same shortcut also rejects all suggested changes.
- Sending a long essay as a correction. One clear fact is easier to act on mid-task.
- Repeating a correction instead of writing a rule, so the same mistake comes back in the next chat.
Sources
- Cursor Agent, Cursor Docs
- Keyboard Shortcuts, Cursor Docs
- Reviewing and testing code, Cursor Learn
- Agent mode, Cursor Help
Checked 6 November 2026. Products change often, so confirm details with the vendor.