Drive Cursor's Agent With a Plan and a Checkpoint

Running Cursor's agent by having it propose a plan for a scoped task, checkpointing to approve that plan, and only then letting it execute across files — rather than turning it loose on a vague goal.

An agent that edits many files is powerful and easy to lose control of. A plan-then-approve checkpoint catches a wrong direction while it's still cheap, and keeps you the reviewer of intent rather than the cleaner-up of a sprawling wrong change.

Multi-step tasks that span files — a feature, a migration, a systematic fix. Best when the task is well-defined enough to plan but big enough to want automation.

A single obvious edit, where planning ceremony is slower than just making the change inline.

Task for agent mode: <scoped goal>. First, produce a step-by-step plan and the list of files you'll touch — do not edit yet. I'll approve or adjust the plan, then you execute and report what changed at the end.

Asked to add input validation across five endpoints, the agent lists the plan and files; you drop one endpoint from scope, approve, and it executes exactly that — no surprise edits.

Give the agent a clear done-condition (e.g. 'all five endpoints reject empty bodies and the tests pass') so it can self-check before handing back, and pair it with scoped context so it stays on target.

• Handing the agent a vague goal and reviewing a huge diff after the fact.
• Skipping the plan checkpoint, so a wrong assumption spreads across every file.
• Never giving a done-condition, so 'finished' is whatever the agent decides.