I listened to Kim Polese's talk at OSCON 2005.
Speaking for SpikeSource, her company, she made the point that in the future, keeping software systems healthy will not just be a big issue, but a strategic one.
Using Open Source Software has become a critical strategy for companies to achieve lower IT costs. But all the different modules have to work together. Plus, they upgrade their versions in very different cycles. Administration and, even more important, testing of these babel-like software hunks will be a big task and if it doesn't become managable, progress in this area will come to a halt.
So far Kim Polese. From my Cognitive Science standpoint, I immediately thought: "This testing - that's a task for... Agents."
Instead of writing code with direct calls to the software like JUnit does, we should abstract these calls somehow from a layer that just describes use cases. Each use case becomes an agent who will act just like the human that described that use case (users, tech support,...).
And each night, hundreds or thousands of little agents test the software for the company.
There is an AI thing to it (the agents: their interface, their messages and so on) but also a plain IT thing (the abstraction of calls to the software babel). And the latter might be the most work. I get the feeling that a lot of today's problems are structured like that.