Danny Ayers has a great post on Atom as Agent Language. I wanted to leave a short version of this post as a comment on his site, but was having trouble contacting his server. So, it is all here.

This follows on some of the ideas I originally posted in Autodiscovery: spider food, interactive query, processing instruction?

I think that, you can start with the assumption that the RSS / Atom type "feed" interaction is more a computer-computer interaction than a human-computer interaction, and also more a peer-peer interaction than server-client interaction. So, if you think of it this way as an interaction of agents, it becomes easier to imagine allowing a richer and more varied interaction syntax and protocol than RSS and Atom.

In other words, I think a lot of the emphasis on the XML syntax (and, implicitly, its data structure) and the serve-a-file-via-HTTP protocol comes from the legacy idea that a person might create such an XML file by hand, and put it, as a static file, on a webserver that can only serve a static file.

Similarly, I think specifics of the XML syntax for RSS and Atom come out of the legacy idea that a feed reader is a client-only polling application, and a feed server (e.g., a webserver) is a server-only application (even largely dedicated to simply serving static files).

This syntax and protocol is a wise and practically required simplicity that was built-in to RSS, but, I think in some interesting cases, it is also a legacy burden on the kind of "smart" computer-computer / peer-peer aggregation scenarios a lot of us are exploring. That legacy interaction is a subset of a larger sphere of interactions for which I imagine a richer syntax and range of protocols that can also appear as a simple RSS / Atom feed / API when it needs to.

(Note: I realize the some people might create RSS or Atom files by hand, and place them as static files on webservers. I am just saying that maybe some interactions don't need to be so fully backwards compatible with this legacy.)