Agent

Agent mode is where Cursor can actively perform multi-step work: exploring the repo, proposing changes, editing files, and running commands.

When to use

• Multi-step work: refactors, feature implementation, cross-file updates
• Mechanical changes that benefit from execution discipline (rename, move, repeated edits)
• When you can validate outcomes with build/tests/lints

When not to use

• Sensitive information is involved (credentials, customer data, restricted client data)
• The scope is unclear, acceptance criteria are missing, or requirements are volatile
• You can’t safely tolerate side effects (risk of destructive commands)

Useful guide

Bound the scope

• Goal and non-goals
• File/folder boundaries (where changes are allowed)
• Constraints (no behavior change, performance, compatibility)
• Definition of done (tests, lint, build, smoke check)

Use small slices

• Prefer small commits/PRs over “mega diffs”
• Separate formatting from behavior changes
• Stop after each slice for review

Verify like normal engineering

• Run the same verification you would require for any PR
• Review diffs like a teammate’s work (look for domain mistakes)

Safety and confidentiality

• Never paste tokens, keys, credentials, customer/production data.
• Prefer minimal snippets; sanitize/redact where possible.
• If in doubt, escalate to your Team Leader / Reporting Manager.