What Value Subagents?

Link post

Among rationalist thinkers the subagent theory of psyche is popular: CFAR teaches a version of internal family systems (IFS), Max Harms writes AI fiction from the point of view of a subagent, Brienne describes herself as metaphorically made up of multiple people, Alicorn seems to have been the first rationalist to explicitly suggest the idea, the notion goes back to our intellectual forebearers, and nearly everyone seems to use a subagent interpretation of dual process theory at the least. Yet although I’ve previously explored multi-part theories of psyche, I don’t think of these parts as subagents. Instead, as I’ve hinted at, I think of the psyche primarily as a unified optimization process with inconsistent preferences that does not decompose into subagents. I didn’t always think this way, though, and have at times thought of myself and others as made up of different configurations of subagents, so I want to explore some of how I came to give up subagent theories and see what value they may still hold.

Stepping back to start, what do rationalists mean by subagents of the psyche? “Agent” has roughly the economic sense of a process that makes decisions, but we should understand “decision” to include any kind of behavior, including thoughts, so in the context of theories of psyche “agent” is a rough synonym for “person” or “mind”. A subagent, then, is an agent operating inside of another agent, forming a system of agents that externally present as a single agent. Explaining how subagents work in IFS should make this clearer.

IFS is a psychoanalytic model that, to my reading, originates from thinking of the Freudian id, ego, and superego like three homunculi competing to control the mind, but replaces the id, ego, and superego with groups of subagents serving exile, firefighter, and manager roles, respectively. These roles differer in both subtle and obvious ways from their inspirational forms, but the basic idea is that exiled subagents are those that would create undesired thoughts and behaviors, managers are subagents that work to keep undesired thoughts and behaviors from occurring, and firefighters are subagents that work to mitigate suffering when exiled subagents gain enough strength to influence a person’s behavior. The psyche can then be thought of like a walled city where managers act as sentries to keep the exiles out and, when exiles inevitably make it inside, the firefighters serve as the garrison that chases the exiles out.

Rationalist thinkers have mostly dropped the Freudian shape of IFS, but they keep the idea of the psyche as a family of subagents. Rather than specifically exiled, firefighting, and managing subagents, the subagents are more often associated with observed aspects of an individual person’s psyche and given attribute names like Lazy, Ambition, Face, and Pride. This internal family is frequently, when first identified, dysfunctional and fighting, corresponding to feelings of cognitive dissonance and conflicting desires. Through the application of various trainable techniques, though, the psychic family can be made functional and happy. Many people I know have expressed finding great value in this model, and they are living more satisfying lives through understanding themselves in this way.

I’ve, however, never much felt like I was made up of many subagents, and when I was introduced to the idea it seemed to me insufficiently parsimonious because it seems to suppose the existence of neurological subprocess different from those that appear to exist. Instead for most of my life I’ve felt like my “I” — my psyche — was a homunculus trying to control a vast, semi-autonomous machine. This is to say my experience matches the description of dual process theory, sometimes referred to in this way as the conscious and subconscious, System 1 and System 2, or as the rider and elephant. The rider is an inner process that is the true self (Dasein if I’m generous; the authentic self if I’m not) and the elephant is an outer process that interacts with the world. The elephant is able to act on its own and has its own ideas about what to do, though, so the rider must make an active effort to direct the elephant. Dual process theory has some biological and evolutionary support, but even in the absence of that it’s correlated with construal level theory, so this seems a potentially useful model.

But thinking of yourself as a rider and elephant can be highly dissociative. I often felt like I wasn’t in control of myself and that I depleted my ego to do what I wanted. I thought of myself as suffering from akrasia because I could exhaust a lot of willpower trying to direct the elephant to little effect. And I lived like this from circa age 10 until I was 30. But as I transitioned from thinking in terms of systems to system relationships, the separation I felt between rider and elephant dissolved. Rather than being made up of two parts, it felt more and more like I was made up of just one. I could still see the shapes of the rider and the elephant, but now I could understand their interactions. This eventually led me to a holonic interpretation of the psyche where there are no distinct parts, but, as we’ll see, there are many details.

That’s because the holon of my psyche can’t have no stuff in it: I know too much about how the brain works to really think of it as a black box. But I can understand it as a box full of gears, where the gears are preferences that interlock, sometimes harmonizing and sometimes grinding, and their combined motions produce my actual behavior. Being a holon, it’s hardly clear exactly how the gears interact, but we have enough clues about the arrangement that it’s possible to use it to make some predictions, not so strong that they risk cracking the epicycles and not so weak that we can claim anything. Unfortunately this theory makes heavy demands for ontological complexity when applied, thus it’s likely not that useful to most folks seeking personal development.

Nonetheless, it’s given me insight into the value others find in many-subagent theories of psyche when viewed through the lens of ego states or “masks”. In Impro, Keith Johnstone describes the theatrical tradition of mask work where performers don masks and “become” the character a mask embodies, sometimes entering a trance where they forget themselves and effectively are the mask’s character. When the psyche is viewed as a box of preferences this has a straightforward interpretation: the mask is a filter for expressing a subset of preferences. And if that’s the case, then we can see psychic subagents as masks that reify subsets of preferences into agents.

In this light subagent theories of psyche make a lot of sense. Rather than being made up of many subagents, each subagent is a mask that allows a person to make salient their preferences that are normally lost in the cacophony of competing desires. By switching between masks you can give voice to the parts of yourself that are often ignored and even let the subagents talk to each other. But because these masks apply an ontology on top of a simpler system that can explain itself without them, subagents truly are like epicycles: they produce correct predictions, they are easier to work with than the full calculus of preference mechanics, and do this despite not describing reality parsimoniously.

So although I never did and no longer need to think of myself as made up of many subagents, I can understand them as an extremely useful form of play. My desire is then that they act as a bridge for people to develop more complete understanding of themselves.