Unfortunately, I think such a thing would be doomed, except on the fringes. You might get a short, profitable run with such a system, but once it becomes noticeably profitable to do so, why wouldn't the service providers do it themselves? If you're able to make money after paying them, they can get what they were paid plus what you made by doing it themselves. And since the APIs must exist for the bots to exist, doing so should be relatively trivial for the providers.
StorJ sounds like more fun. A similar bot which makes money so it can pay humans to do its bidding (help it reproduce), which means it doesn't require widespread APIs that would contribute to its own downfall.
Some similar musings from history: http://www.research.ibm.com/massive/bump.html http://www.cap-lore.com/Agorics/Library/agoricpapers.html
Not necessarily, you could have a self modifying agent using genetic algorithms that could develop quicker than having a coder tweaking everything.
You can't solve chess with a genetic algorithm, at least not to my knowledge, but you could sure as heck make a challenging checkers player with one.
And yes, the designs the algorithm come up with can in some ways be considered superior to what a human can come up with.
These kind of agents would have at least some competitive advantage over humans in providing services which are considered illegal (for example spamming). Humans probably need to price in the risk of getting caught and ending up in jail while an agent does not have the problem.
However, I think that the coupling TC/virtualization is still at an early stage of development and I'm not aware of any cloud infrastructure providing virtual TPMs.