As a pre-paradigmatic field, one of the main goals of agent foundations is to become less confused. Agent foundations researchers have strong intuitions that certain formal objects should be definable, and that they should have certain definable or provable properties. But this work hasn’t succeeded yet because we are still very confused about how to properly think about it.
Research towards deconfusion is often contrasted with research towards practical application. It can sometimes be hard to see how fundamental research could ever be applied, because it happens at a very general level. Agent foundations researchers tend to believe that robust solutions to the applications will require deeper deconfusion before they can be concretized towards specific applications. For more detail on this through-line, see how theory can help AI safety.
