Yeah, I’m not aware of any AGP cards (though there is the regular PCI card mentioned).
Part of the reason for this is bus speed. CUDA depends a lot on how fast you can move data between the GPU and CPU. PCI on consumer motherboards tops out at 133 MB/sec nominal, while AGP in its final form hit just over 2 GB/sec nominal. PCI Express started at 4 GB (for an x16 slot that a graphics card goes into) nominal, which in practices reaches 3.2 GB/sec in CUDA. Now it’s double that, and next year it will double again.
The other aspect is marketing: The gamer audience that fuels much of the high end card sales were lining up to buy these cards because of DirectX 10 support. (DX10 support in NVIDIA cards came at the same time as CUDA) That demographic is more than willing to upgrade to a newer motherboard with a nice bus to support their new video card. This means that there is little demand to build a new GPU chip with a native AGP interface, so building an AGP card would require some kind of translator bridge chip. That seems to be both hard and expensive.