I'll preface the following by saying that I haven't actually played the new Sim City yet, but I do plan to pick it up sometime soon (once the server issues are ironed out). I do, however, have some small amount of knowledge of academic economic simulation, so hopefully my input might be worth something.
Firstly, agent-based modeling of economic systems is insanely fucking difficult. Even ignoring the computational restrictions, it's a seriously difficult problem to figure out, and there are plenty of very intelligent people trying to do so.
For those that are working on it, they generally only do so with very low-level simulations, modeling a small number of people making simple, isolated decisions. Even then, and without any computational limitations, it's a huge challenge to create a simulation that produces the expected behavior both on the individual level and on the aggregate level.
Maxis, though, are not only attempting to find solutions to these problems, they're doing so on a larger scale and with more interconnected systems than any academic simulation I've ever heard of. The fact that they've managed to create something of this scale that works at all is absolutely astonishing as far as I'm concerned.
Then we move onto the computational limits. If a group of academics were to work on an agent-based simulation of the scale of Sim City, they wouldn't attempt it on anything less than thousands of cores worth of supercomputer, and wouldn't even attemp to run it in real-time even then. Maxis, however, have managed to do so in a real-time game that runs (afaik) on a Core 2 Duo. Again, the fact that it runs at all is a serious achievement.
When it comes to path-finding for vehicles and the like, the big limit that Maxis are coming up against is the computational power of the average PC. Finding the optimal path over a non-Euclidean graph in minimal computational time is something that has a known and relatively simple solution. The issue is that when you're dealing with a graph as complex as the road network in one of Sim City's cities, and are trying to run the path-finding algorithm for hundreds or thousands of cars simultaneously, you need orders of magnitude more computational power than a home PC is going to provide. The only option you then have is to move to a non-optimal heuristic, which finds a "good guess" as to the optimal route. This is what 99% of games do when it comes to pathfinding, as it's computationally much cheaper and for the most part it produces paths that look pretty good. The issue, though, is that any non-optimal pathfinding heuristic is, by definition, going to throw up imperfect routes every so often. This is an even bigger issue when it comes to routing around traffic in a game like Sim City, as you're then dealing with non-Euclidean graphs, which adds a whole extra level of complexity.
There's simply no way around this when working within the confines of a typical PC. Either you go with an optimal solution and the game starts to chug with more than a couple of cars on the road, or you use a heuristic and end up with the occasional illogical traffic jam. It's a mathematically proveable fact, and there's nothing that Maxis can really do about it.