Personal Modeling Through Models and Harnesses

Contact Me

In Two Signals for Harness RSI, I discussed how a personal agent can learn from a user’s instructions, corrections, and choices. These signals help the harness adapt to how that person wants to work. A correct result becomes more useful when it also fits the user’s goals, preferences, and circumstances.

This suggests a broader direction. As an agent learns to serve a person, it accumulates a representation of that person. The model supplies general capabilities; the harness supplies personal context and procedures for using those capabilities. Together, they can become an executable model of how the person might choose, respond, and work in a particular situation.

Personal modeling could connect individual agents to population simulation. If we can build useful models of many people, we can explore their different needs, study demand, and investigate how work might be organized. Recommendation, advertising, and labor modeling become applications of the same underlying question: what would these people do under these conditions?

A Personal Harness Fits a Person Across Domains

A domain harness organizes an agent around a kind of work. A coding harness might provide repository tools, validation procedures, and review workflows. A personal harness has another organizing principle: the person it serves. That person’s priorities continue across writing, shopping, planning, learning, and professional work.

Consider someone who values control over their schedule, dislikes maintaining complex tools, and prefers paying more when it reliably saves time. Those priorities could affect which software they adopt, which trip they book, and how they delegate a project. Each domain needs its own knowledge and tools; the personal harness helps apply them according to the same person’s context.

This makes personal fit a central objective for a personal agent. General competence establishes what the system can do. Personal adaptation helps determine which of those capabilities are useful now, which tradeoffs to present, and when to ask the user to decide.

The harness RSI loop can sustain that fit. It observes mismatches, proposes changes to memory or workflow, and evaluates whether later interactions improve. Its target includes both changing technical dependencies and a person whose circumstances and priorities also change.

The Person Model Lives in the Whole Agent

Personal modeling does not require training separate model weights for every person. A shared foundation model can interpret different personal memories, instructions, and histories through different harnesses.

A useful conceptual representation is:

shared model capabilities
          +
personal harness
  memories and supporting evidence
  goals, preferences, and constraints
  tools and working procedures
          +
current situation and available choices
          |
          v
predicted response or proposed action

The representation is distributed. A memory may record that the user avoids overnight travel. A planning procedure may prioritize uninterrupted work time. A tool may expose the calendar constraints that make an otherwise attractive option impossible. The model brings these pieces together when considering a new situation.

This is more expressive than a list of interests. It can represent conditional tradeoffs: a person may prefer the cheapest option for routine purchases but pay for reliability when a deadline matters. A useful model preserves the conditions under which a preference applies and the evidence supporting it.

There is also a distinction between helping someone and predicting them. An assistant may recommend a new habit that would help its user, while a simulator tries to predict whether the person would actually adopt it. These purposes need explicit objectives and separate evaluation, even when they use the same personal context.

From Personal Models to Virtual Populations

Once a person model can respond to scenarios, we can assemble many such models into a virtual population. We could present a new service, vary its price or features, and examine how responses differ across people and circumstances. The result could expose competing needs that an average preference hides.

There is research evidence for parts of this direction. LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals built agents using interviews, surveys, or both from 1,052 Americans and evaluated predictions on held-out survey items and other behavioral measures. This supports investigating agents grounded in evidence about individuals. The continuing personal harness proposed here extends that direction toward learning through ongoing assistance.

Population modeling introduces another layer. A collection of independent responses can estimate a distribution of preferences under a fixed scenario. Modeling adoption over time also requires assumptions about how people influence each other, share resources, and encounter changing prices or availability. A product that looks attractive to every isolated agent may become less useful when everyone competes for the same limited service.

Scale therefore involves both the number of represented people and the environment connecting them. Copying one agent many times adds simulated responses without adding evidence about additional people. A population needs coverage of the relevant differences and an explicit account of who is missing.

Recommendation and Advertising Through Personal Simulation

Simulation already has a place in recommender research. Google’s RecSim provides configurable environments for sequential recommendation, including assumptions about user preferences, changing states, and response behavior. Personal models suggest a way to investigate richer user representations within this broader simulation approach.

For example, imagine a service for learning a professional skill. A virtual population could compare a cheap course requiring a fixed weekly schedule, a flexible course with less support, and a more expensive course with coaching. The simulation could explore which constraints drive rejection, which groups remain underserved, and which product changes are worth testing with people.

This could move recommendation toward evaluating how an option fits a person’s current goals and constraints. Advertising could use related simulations to explore whether an offer addresses a need, which information helps people assess it, and how interest changes with price or timing. These are hypotheses for real experiments; simulated interest alone does not establish demand or incremental sales.

The objective matters. A personal agent chooses on behalf of its user. An advertising system may optimize a seller’s return. The same representation can produce different decisions under those objectives. Sharing a personal model with another system should therefore be a separate, user-controlled decision, with a defined purpose and limits on the information exposed.

One possible design keeps the detailed personal context with the agent. An external service submits an offer, and the agent evaluates it locally against the user’s needs. Population analysis could use permitted aggregate results. The access boundary becomes part of the system design rather than assuming that personalization requires exporting a person’s history.

Understanding Needs and Modeling Work

The same approach could explore needs before a product exists. Suppose many agents struggle to find a service satisfying their users’ scheduling, accessibility, or budget constraints. Those repeated mismatches suggest demand that the current catalog does not meet. A population simulation could help formulate and compare possible services before a team invests in building one.

Work adds the supply side. Personal models could represent aspects of a worker’s skills, availability, preferred tasks, coordination habits, and willingness to accept different assignments. Combined with a task environment, they could help explore staffing, training, and collaboration between people and agents.

For example, a team considering an AI tool could simulate several ways of distributing research, implementation, and review. Different scenarios might reveal where supervision becomes a bottleneck or where a new skill would make someone more effective. The output would be a set of assumptions and predicted outcomes to test in actual work.

Predicting a person’s preferences does not establish their productive capacity. A simulated worker needs evidence about task completion, quality, time, and coordination under relevant conditions. An agent that writes a convincing account of someone’s work has not demonstrated that it can perform that work. Using an agent to carry out tasks also requires its own tools, access, and execution checks.

With those distinctions, personal modeling offers a way to study demand and work together: what people need, what they are willing and able to do, and how new tools might change the relationship between the two.

Validate the Model Against the Person

The central test is whether the model predicts useful aspects of its person’s behavior on new situations. A plausible explanation is easy to mistake for understanding. Evaluation needs choices and outcomes that were unavailable when the relevant personal representation was built.

For individual models, compare predictions with later decisions and track where the model is uncertain. For populations, compare predicted distributions with observed responses, including differences between groups. Check whether personal context improves predictions over a shared model given only the scenario or broad demographic information.

Models sharing the same foundation model may also share errors. More simulated agents cannot remove a common mistaken assumption. Real observations are needed to identify those errors and revise the representation or environment.

Personal models must remain correctable as people change. A person should be able to inspect, revise, or remove the memories and inferences used to represent them. Predictions should retain their context and uncertainty, especially when they inform decisions about opportunities or work.

Personal Adaptation as a Foundation for Population Modeling

The immediate value of a personal harness is better assistance for one person. Ongoing interaction gives it opportunities to learn which goals, constraints, and working methods matter, then test that understanding through later work.

That creates a path toward a broader capability:

personal assistance and feedback
              |
              v
model and harness adapted to an individual
              |
              v
validated predictions in new situations
              |
              v
populations with explicit environments and interactions
              |
              v
experiments on demand, recommendations, and work
              |
              v
real outcomes that refine the models

The opportunity is to make accumulated personal understanding usable across scenarios. A harness that increasingly fits one user can become a foundation for modeling that person’s choices. Many validated person models, connected through an appropriate environment, could help us understand groups and design services and ways of working around their different needs.