This post muses on a potential computational framework for psychopathology, built around two tentative ideas:

  • Internal Agents: The mind could perhaps be viewed as a system of interacting “agents” (e.g., a “self-critic,” “problem-solver”) that seem to influence mood and behavior.
  • External Events: Life stressors might be conceptualized as probabilistic streams, reflecting their likelihood of happening over time.

Connecting these concepts could suggest a lens for thinking about intervention, pointing toward possible ways to either address external factors or foster more adaptive internal responses.

I had a curious thought cross my mind today, a sort of conceptual model for thinking about the dynamic nature of our internal worlds. It’s a way to bridge systems thinking, psychopathology, and some of the newer tools in computational modeling. This idea is in it’s infancy, and this represents my first attemp to get down the basics.

The idea begins with a framework I often come back to: modeling psychological distress, like suicidality, as a network of interacting factors. We often talk about internalizing factors (like rumination or negative self-talk) and externalizing factors (like impulsivity or aggression). We can visualize these as nodes in a directed acyclic graph (DAG), showing how one symptom can trigger another.

But here’s the thought that struck me as new: what if we modeled these internal factors not just as static nodes, but as agents?

The Internal System: A Society of Agents

Imagine the internal world as a collection of different “parts” or “voices,” a concept seen in therapeutic approaches like Internal Family Systems Therapy. It’s the idea that within us, we have different roles that take the lead at different times. There might be a “social planner,” a “harsh critic,” an “anxious protector,” or even a “wise advocate.”

Now, what if each of these parts was a computational agent, maybe even powered by a large language model? Each agent would have its own function, its own level of competence, and its own triggers for activation. When faced with a challenge, the system would “call upon” different agents to help regulate emotion or plan a response.

A system like this would be incredibly dynamic. The overall mood or behavior of the person wouldn’t be a simple sum of its parts, but an emergent property of the interactions between these agents. You could have a highly competent “internal clinical psychologist” agent that suggests helpful behavioral strategies, but it might be rarely activated because a louder, more reactive “anxious protector” agent always jumps in first.

This moves beyond thinking about intervention as simply a list of behavioral techniques. It frames it in a more dynamic, language-based way. How do we help the system learn to call on its more adaptive agents? How do we bolster the skills of a struggling agent?

The External World: A Probabilistic Stream of Events

Of course, our internal system doesn’t operate in a vacuum. It’s constantly being influenced by the outside world. This is the second piece of the puzzle: modeling contextual factors.

Life events—a negative comment from a boss, a fight with a partner, getting good grades, a delayed train—are inputs that can either regulate or dysregulate our internal system. The interesting part is that while we can’t predict exactly when a specific negative event will happen, we can often model its general rate of occurrence.

A tool like a Poisson distribution immediately comes to mind, and it may be quite suitable for this. It’s used to model the probability of a given number of events happening in a fixed interval of time, assuming these events occur independently and at a constant average rate. So, for a person experiencing chronic social friction, their “negative social event” category would likely have a high rate of occurrence—a high lambda ($\lambda$) value, in statistical terms. This mathematically captures the idea of a continuous “stress load” that bombards the individual’s internal system. One could imagine having different probabilistic models for different categories of life events: work, social, health, etc. This would need to be considered more carefully, obviously.

Bringing It All Together in a Unified Framework

This is where the directed acyclic graph comes back in. A DAG could visualize the entire system:

  1. External Event Nodes: These would be the probabilistic generators of life stressors, like “Academic Stress” or “Relationship Conflict,” each with its own rate of occurrence modeled by a Poisson distribution.
  2. Internal Agent Nodes: These are the internal voices, like the “Self-Critic,” “Social Planner,” or “Problem-Solver.”
  3. State Nodes: These could represent mood, anxiety levels, or energy.

The edges of the graph would show the influence. An event from the “Academic Stress” node (e.g., a bad grade) might activate the “Self-Critic” agent. The “Self-Critic” agent, in turn, would strongly influence the “Mood” state node in a negative direction. Meanwhile, a positive event might activate a different, more helpful agent.

A Conceptual Path to Intervention

Thinking about this as a case formulation is what really excites me. It makes the potential targets for change so clear. You could intervene in two fundamental ways:

  1. Treat the External Nodes: This involves addressing the contextual factors themselves. If the “Delayed Train” node is a major source of dysregulation, a practical intervention might be exploring alternative transportation. It’s a direct attempt to lower the rate of a specific stressor.
  2. Treat the Internal Nodes: This is about changing the internal system’s response. It could involve “bolstering” the helpful agents—making the “internal psychologist” more skilled and more easily activated—or helping the system learn to down-regulate the unhelpful agents. This feels like a powerful computational metaphor for the work of psychotherapy.

Ultimately, this is just a sketch. But I like how it brings together the internal, agentic self and the probabilistic nature of the external world into one dynamic system. It helps conceptualize a person not as a static collection of symptoms, but as a complex, self-regulating entity constantly adapting to a stream of life events. It’s a compelling way to think.