Agents are not average tools. They are smart and resourceful, so they may bypass instructions if they see more efficient options, or they may simply make mistakes, which happens with intelligence. Therefore, in my experience, setting boundaries tends to work well. They may make mistakes, which are recoverable if failures are taken into consideration and guardrails are set, but they learn the lessons much more quickly. Their agency, which is annoying sometimes, can also keep design honest and efficient.