Workflows
Steer Agent with a plan and checkpoints, redirect it mid-run, set what it may run alone, and review before you commit.
Agent loop
Plan first, use checkpoints, redirect Agent while it works, choose a Run Mode, and run Agent Review before you commit.
4 techniques →-
Which Run Mode Should You Use in Cursor? Run Modes decide when Agent can run shell commands, MCP tools and fetches without asking you. Level 3: Workflows
- 1Auto-reviewallowlist, then sandbox, then classifier
- 2Allowlistyour list runs, the rest asks
- 3Run Everythingno prompts, no sandbox
-
How Do You Keep Cursor's Agent on Track With a Plan and Checkpoints? Press Shift+Tab in the chat input until the mode shows Plan, describe the task, answer its questions, edit the plan, and click Build. Level 3: Workflows
- 1Clarify and research
- 2Plan
- 3You edit it
- 4Build
- 5Revert, refine, rerun
Then repeat -
How Do You Redirect Cursor's Agent While It's Working? You don't have to wait for Agent to finish. Level 2: ContextAgent keeps goingAgent stops
- Queue (Enter)
- Send now (Cmd+Enter)
- Stop
-
How Do You Get a Code Review From Cursor Before You Commit? Agent Review is a dedicated code review on your local changes. Level 3: WorkflowsAgent Reviewbefore you commit
- Reviews local changes
- Quick for small diffs
- Deep for sensitive code
Bugboton the pull request- Reviews on push
- Starts at Premium
- Ignores .cursor/rules
Can you steer Cursor's agent?
Three questions on approvals, plans and redirecting a running agent. See your answer explained right away, with no sign-up.
You choose how much the agent decides alone
Working with Agent involves a few controls that most people leave at their defaults. Plan mode decides when you approve the approach. Run Modes decide which commands need your OK. Queuing and steering decide how you step in during a run. Agent Review adds a check at the end. Setting these deliberately matters more as your tasks get longer.
A run that goes wrong is normal
Cursor's docs recommend something that feels backward: when a build misses, don't patch it with follow-ups. Revert, make the plan more specific, and run it again. Checkpoints make this cheap, and it often produces cleaner results than repairing an approach that started in the wrong direction.
What is the Cursor agent loop?
Agent reads code and uses tools to edit files and run commands, then checks the results and continues until the task is done. Cursor's docs say there's no limit on how many tool calls it can make during a task.
How do I stop the Cursor agent from making too many changes?
Use Plan mode for wide tasks so you can see the file list first, state what's out of scope, and keep each request small. Click Stop if you see it heading the wrong way.
Can Cursor run commands without asking me?
That depends on your Run Mode, under Settings, Agents, Approvals & Execution. Auto-review runs allowlisted calls, sandboxes other shell commands when it can, and uses a classifier for the rest. Cursor's docs say it isn't a security boundary.
How do I make the Cursor agent produce better plans?
Give it a clear goal, the scope, a pattern to follow and a done condition, then answer its questions and edit the plan before clicking Build.