We study a new optimal stopping problem: Let $G$ be a fixed graph with $n$ vertices which become active on-line in time, one by another, in a random order. The active part of $G$ is the subgraph induced by the active vertices. Find a stopping algorithm that maximizes the expected number of connected components of the active part of $G$. We prove that if $G$ is a $k$-tree, then there is no asymptotically better algorithm than `wait until $\frac{1}{k+1}$ fraction of vertices'. The maximum expected number of connected components equals to $$\left(\frac{k^k}{(k+1)^{k+1}}+o(1)\right)n.$$