A try: An Agent is any object (which presumes a mechanism which clusters perceptions into objects, but which is a more general and separately solvable problem) whose interactions with other objects can best be modeled (in the sense of prediction) as having an inner model of the environment (which doesn't require recursion but doesn't preclude it).
I would be very grateful for as many different attempts at rigorous or semi-rigorous definitions of an “agent” as possible.
Specifically, a definition of an (embedded) Agent that makes it intuitively clear the nature of the boundary between Agent and Environment.
(I have read up on Agency, Reductive Agency, Causality, True Names, Natural Abstraction, Boundries, FEP… you get the idea.)
I think I have some intuition of what a formal description of an “Agent” could look like qualitatively - but I think contrasting this with other perspectives would quickly draw out any overlooked failure modes.
If you have a sense of why you chose your “lowest level” to be your lowest level, and not further reducing the smallest components (e.g. describing the boundary all the way down to causality in quantum fields, in an extreme case), please do add that as well!