One possible change here is that agents in the 90s didn't really have anything to "agent". There's a lot more useful APIs now that might allow a much better cost/benefit tradeoff in creating an "agent" that doesn't require so much raw screen scraping and the associated (staggering and really easy to grossly underestimate) upkeep.
Like much of the rest of the 1990s Internet bubble, arguably the problem isn't that the ideas were all bad, but that they were running well in excess of what the infrastructure could support. Even the much-maligned Pets.com was really only a bad idea for this reason. Pet food delivery went from crazy-stupid idea to mundanity without much fanfare: http://www.amazon.com/Iams-Proactive-Health-Control-17-5-Pou... . Well, that and the silliness of the idea that you could or would build that infrastructure and then only use it to deliver pet-related things, rather than build an Amazon.
Yes, there were some very early implementations on mainframes and research computers, but the first widely-deployed spreadsheet was VisiCalc, released in 1979, and it really wasn't until the 1980s that PCs became remotely commonplace in commercial offices. Arguably not particularly useful until the early 1990s with Windows 3.1, though there were some pioneers ahead of that time.
