In my recent post entitled The MSDN Camp vs. The Raymond Chen Camp I wrote

Our team [and myself directly] has gone through a process of rethinking a number of decisions we made in this light. Up until very recently we were planning to ship the System.Xml.XPath.XPathDocument class as a replacement for the System.Xml.XmlDocument class
The problem was that the XPathDocument had a radically different programming model than the XmlDocument meaning that anyone who'd written code using the XmlDocument against our v1.0/v1.1 bits would have to radically rewrite their code to get performance improvements and new features. Additionally any developers migrating to the .NET Framework from native code (MSXML) or from the Java world would already be familiar with the XML DOM API but not the cursor-based model used by the XPathDocument. This was really an untenable situation. For this reason we've reverted the XPathDocument to what it was in v1.1 while new functionality and perf improvements will be made to the XmlDocument. Similarly we will keep the new and improved XPathEditableNavigator XPathNavigator class which will be the API for programming against XML data sources where one wants to abstract away what the underlying store actually is. We've shown the power of this model with examples such as the ObjectXPathNavigator and the DataSetNavigator.

I've seen some concerned statements about this posts from XML developers who use System.Xml such as Oleg Tkachenko, Fumiaki Yoshimatsu and Tomas Restrepo so it seems I should clarify some of the decisions we made and why we made them.

In version 1.0 of the .NET Framework we provided two primary classes for interacting with XML; the XmlDocument and XmlReader. The XmlReader provided an abstract interface for interacting with a stream of XML. One can create an XmlReader over textual XML using the XmlTextReader or over virtual XML data sources such as is done with the XmlCsvReader. On the other hand with the XmlDocument we decided to eschew the approach favored by the Java world which used interfaces. Instead we created a single concrete implementation. This turned out to be a bad idea. It tied the the interface for programming against XML in a random access manner with a concrete implementation of an XML store. This made it difficult for developers who wanted to expose their data sources as XML stores and led to inefficient solutions such as the XmlDataDocument.

To rectify this we needed to separate the programming model for accessing XML data sources from our concrete implementation of the XmlDocument. We chose to do this by extending the cursor based programming model we introduced in v1 with the XPathNavigator instead of moving to an interface based approach with XmlDocument. The reason for choosing to go with a cursor based model over a tree based model is summed up in this quote from my article Can One Size Fit All?

In A Survey of APIs and Techniques for Processing XML, I pointed out that cursor-model APIs could be used to traverse in-memory XML documents just as well as tree-model APIs. Cursor-model APIs have an added advantage over tree-model APIs in that an XML cursor need not require the heavyweight interface of a traditional tree-model API where every significant token in the underlying XML must map to an object.

So in Whidbey, the XPathNavigator will be the programming model for working with XML data sources when one wants to abstract away from the underlying source. The XPathNavigator will be changed from the v1.0 model in the following ways (i) it will be editable and (ii) it will expose the post schema validation infoset. I've already worked with Krzysztof Cwalina on updating the Design Guidelines for Exposing XML data in WinFX to account for this change in affairs.

As for the XPathDocument, it is what it always has been. A class optimized for use in XPath and XSLT. If you need 10% - 25% better perf [depending on your scenario] when running XPath over an XML document or running XSLT over in-memory XML then this class should be preferred to the XmlDocument.