Hand Off Scoped Tasks to Background Agents

A workflow for Cursor's agents: give a well-scoped task with clear done-criteria and let the agent plan, edit across files, and run commands — while you review the result as a diff.

Agents are strongest on bounded, verifiable work and weakest on open-ended 'make it better' asks. A clear scope plus a way to check success (tests pass, a script runs) lets the agent self-correct, and turns your job into reviewing a diff rather than typing every edit.

Mechanical-but-spread-out tasks: rename a concept across the repo, add a field end-to-end, migrate a deprecated API, wire up a new route following an existing one.

Ambiguous design work or anything where 'done' can't be checked. If you can't state the success condition, the agent can't know when to stop — do that thinking first.

"Task: add a `deletedAt` soft-delete field to the Order model, end to end.
Scope: model, migration, repository, and the two queries that list orders.
Done when: `npm test` passes and listing excludes soft-deleted orders.
Follow the pattern already used for `archivedAt`. Show me a plan before editing."

Asked to migrate 30 call sites from a deprecated fetchUser() to getUser(), the agent finds them via the index, edits each, runs the tests, fixes the two that break, and hands you a clean diff to review — a 40-minute chore done in one pass.

Point the agent at an existing reference implementation ('follow the pattern used for archivedAt') so it matches your conventions, and require a plan before edits so you can correct course cheaply. Combine with strong rules so its output already fits the codebase.

• Giving an unscoped task with no success condition, so the agent wanders and over-edits.
• Not asking for a plan first, then discovering a wrong approach after 20 files changed.
• Rubber-stamping the diff — always review agent output; scope makes it checkable, not infallible.