A staff engineer says his LLM use moved from small assists to starting every code change, bug investigation, test setup, and codebase research with an agent. The grounded part is the review boundary he keeps around it, including rejecting many attempts, writing important PR text himself, and still avoiding agent-led UI judgment.