As others have pointed out, "knocking on people's doors" can be difficult when the said doors are spread over the World, in locations which are geographically remote and, more importantly, legally remote. If all the Tor nodes are under in scope of a single law enforcement agency (e.g. they are all in China, or all in the USA), then that agency may indeed knock on all doors and have them opened. In an international context, this becomes harder.
Further complicating the issue is that "honest" Tor nodes (those who follow the protocol faithfully) do not keep logs, so even if they want to comply to a demand for information on a past connection, they simply do not have it any more.
Breadcrumb following is not the only way to unravel Tor, though. As the metaphor goes, if Hop-o'-My-Thumb had been smarter, he would not have brought bread, but a compass. In the case of Tor, information can be obtained by correlation in time and size. When a Tor connection occurs, the data goes through a randomly chosen sequence of nodes; as the theory goes, as long as one of these nodes is honest, the original client cannot be traced back. Or can he ?
If the evil attacker runs several nodes (and I expect big agencies to actually run a lot of Tor nodes, if only to remain "in the loop"), then it may happen that both the entry node and the exit node are attacker-controlled. These two nodes can then notice the case when a data packet entered the entry node, and a packet of the same size exited the exit node a few milliseconds afterwards. A normal connection entails several successive packets back and forth, further strengthening the correlation. The exit node can see the packet contents (and, at least, their final destination), while the entry node knows the client's identity (well, at least its IP address). That's all that is needed to destroy anonymity -- and that attack works regardless of the honesty of the nodes in between, in flagrant contradiction with the overoptimistic theory alluded-to above.
("One honest node is sufficient" is the theoretical framework for Mix networks as used in electronic voting protocols -- but that works only in the context of voting, where a node can accumulate all packets and send them en masse at the end.)