LLM-driven hardware design remains unreliable because it is largely open-loop, producing syntactically correct RTL without the iterative grounding provided by EDA tools. We refer to this disconnect as the Tool-Feedback Gap, which is further compounded by PPA blindness, where post-synthesis power, performance, and area constraints are ignored. This paper surveys emerging closed-loop, agentic EDA systems that integrate simulation and synthesis, and highlights a growing shift toward multi-agent orchestration with specialized planner, executor, judge, and debug roles to filter noisy tool outputs and mitigate context-switching failures. It further calls for benchmarks that measure convergence efficiency and PPA-aware optimization, rather than simple pass/fail simulation, and warns that autonomous feedback loops can amplify security risks such as vulnerability injection, hardware Trojans, and vacuous verification. These risks motivate the need for standardized tool interfaces, dynamic routing to specialized smaller models, and reasoning-based repair to ensure safe and reliable agentic hardware design.