I made no attempt to explain what an agent would be, only to explain the overall control flow of the data collection process, which to me felt like the weakest part of the paper: the outer algorithm one executes to actually run the algorithms given in discovering agents. Very roughly an agent is a part of the mechanism graph that responds to state forward in the object-level graph and moves it backward in the mechanism graph to influence earlier parts of the next cycle of the object-level graph. I’d love to try to get a more precise explanation written but my understanding is still not high enough quality to do a good job myself. I’ll give it another shot in a bit, or if you’re interested in hopping on a discord call send me a dm here on lesswrong and I’ll give you my discord username.
I made no attempt to explain what an agent would be, only to explain the overall control flow of the data collection process, which to me felt like the weakest part of the paper: the outer algorithm one executes to actually run the algorithms given in discovering agents. Very roughly an agent is a part of the mechanism graph that responds to state forward in the object-level graph and moves it backward in the mechanism graph to influence earlier parts of the next cycle of the object-level graph. I’d love to try to get a more precise explanation written but my understanding is still not high enough quality to do a good job myself. I’ll give it another shot in a bit, or if you’re interested in hopping on a discord call send me a dm here on lesswrong and I’ll give you my discord username.