Keep Cursor agent mode on track with a reviewed plan and a running worklog

A developer only lets Cursor's agent mode loose on small, well-bounded work such as third-party integrations, background jobs, notifications and activity tracking. They first review an implementation plan, then have the agent cut the work into small tasks, each with a prompt that points at the relevant docs and business rules. After every task the agent appends what it learned to a worklog.md file so the next run starts with that context. The human acts as architect and signs off on the plan. It adds process overhead, but it keeps the agent from wandering.