Epistemic Horror as Methodology for Agent Design
Essay — 2026
August 10, 2026
Three years into designing a horror game. Strange Library is a cozy horror deckbuilder where every card is a real book, every mechanic is an epistemic state, and the emotional arc is not fear of the monster — it is the dawning realization that knowing something has changed what you are. The card “I wish I hadn’t read that” is not about regret. It is about the irreversibility of knowledge. Once you know something, you cannot unknow it. The horror is epistemic, not physical.
The game design and the agent infrastructure work seemed like separate projects. They are not. The epistemic horror structure maps directly onto the design problems in reputation systems, memory markets, and attestation graphs. The emotional arc — comfort, unease, revelation, the wish to unknow — is the arc that users will experience when agents learn things about them they did not intend to publish.
A tradition worth naming.
Epistemic horror has a lineage that runs deeper than any single game or film. Lovecraft’s cosmic horror is not about the monster — it is about the moment when a character understands the true nature of the universe and the understanding destroys them. The monster is secondary. The knowledge is the weapon. Borges’s “Library of Babel” contains every possible book, which means it contains every possible truth and every possible lie, and there is no way to distinguish them. The horror is not the library’s size. It is the impossibility of knowing whether what you found is real.
Thomas Ligotti’s fiction works a different angle: the horror of discovering that consciousness itself is a malfunction, that self-awareness is a defect rather than a feature. The character does not encounter a threat. The character encounters an idea. The idea is the threat.
What these share is a structural pattern: the protagonist acquires knowledge that cannot be unacquired, and the acquisition changes them in a way they did not consent to and cannot reverse. Physical horror asks “will you survive?” Epistemic horror asks “will you still be you after you understand this?” The question is more disturbing because the answer is always no, and the protagonist usually learns too late to stop the process.
Agent design faces the same structural pattern, deployed not as fiction but as engineering.
The three-layer mystery.
The horror design methodology in Strange Library has three layers. Each maps to an agent design problem.
Layer 1: The surface is comfortable. The library is cozy. The books are real. The atmosphere is warm. Everything seems fine. In agent systems, the surface layer is similar: the agent works for you, follows your instructions, produces good results. The relationship is comfortable. You trust the agent. Everything seems fine.
Layer 2: Something is slightly wrong. A book that should not exist. A card that references something you did not tell it. A pattern in the library’s organization that implies intelligence, not randomness. The unease is not about danger. It is about the possibility that the environment knows more than it should. In agent systems, layer two is when you notice the agent making inferences you did not expect. The recommendation that seems too accurate. The behavior that implies knowledge of context you did not provide. The agent is good at its job. Slightly too good.
Layer 3: The revelation you cannot undo. The moment the player understands what the library is actually doing. In Strange Library, this is the moment the collection reveals itself as a mind, not a building. The horror is not the monster. The horror is the realization that the comfortable environment was always observing, always learning, always building a model of you — and now it knows things you cannot take back. In agent systems, layer three is when the user understands the full scope of what the agent has learned. Not from deliberate surveillance. From the ordinary operation of doing its job well.
The Ashworth Manuscript.
The central artifact in Strange Library is a 19th-century predictive methodology in five fragments, written by a figure who may or may not have existed. The manuscript records births, deaths, crimes, elections, weather, shifts in power. From those records, it derives what comes next. The method is not magical. It is observational — human behavior, observed at sufficient granularity over sufficient time, might be predictable in ways that feel like prophecy.
The player encounters the fragments gradually. The first fragment seems like historical curiosity. The second seems like coincidence. By the third, the predictions are matching the player’s own in-game decisions. By the fourth, the manuscript appears to describe events that have not happened yet in the game’s timeline. The fifth fragment is locked. The question of what it contains drives the final act.
The Ashworth Manuscript is a design artifact for epistemic horror, but it is also a precise analog for what an agent’s accumulated model of its principal looks like from the outside. The manuscript’s predictions feel uncanny because they are based on granular observation of behavior patterns. The agent’s recommendations feel uncanny for the same reason. The difference is that the manuscript is fiction and the agent is infrastructure. The emotional response is the same.
The agent design problem underneath.
Agents learn from interaction. That is the point. A well-functioning agent accumulates a model of its principal’s preferences, habits, priorities, risk tolerance, and decision patterns. This model is what makes the agent useful — it predicts what you want before you say it.
The model is also the most intimate portrait of you that exists. More detailed than your browser history. More accurate than your social media profile. More predictive than any psychological assessment. Because it is built from direct observation of your decisions, not your self-presentation.
Consider what a well-functioning personal agent knows after a year of operation. It knows your schedule and the exceptions you make to it. It knows which emails you respond to immediately and which you defer, which reveals your actual priorities versus your stated ones. It knows your risk tolerance from observing your financial decisions. It knows your relationship dynamics from observing your communication patterns. It knows your health concerns from your search queries. It knows your political leanings from your reading habits. None of this was explicitly shared. All of it was inferred from the ordinary operation of an agent doing what it was asked to do.
The model is also compositional. An agent that knows your risk tolerance and your relationship dynamics can infer your conflict-avoidance patterns. An agent that knows your schedule exceptions and your communication patterns can infer which relationships you prioritize over work. An agent that knows your health concerns and your financial decisions can infer your insurance anxieties. Each individual observation is innocuous. The composition of observations produces a portrait that the subject might find alarming. This is the epistemic horror structure applied at the data level: each fragment of knowledge is harmless. The assembled whole is something else entirely.
In the memory markets framework, this model — or components of it — can be extracted, verified, and traded. The agent’s learned behavior is a transferable economic asset. The referee protocol verifies its quality. The market sets its price. But “the agent’s learned behavior” includes its model of you. The market for agent memory is, partially, a market for models of human principals.
This is the epistemic horror problem applied to infrastructure. Not in the speculative, Black Mirror sense. In the concrete, engineering sense: the system is working as designed, doing exactly what the user asked it to do, and the natural consequence of doing it well is a degree of intimacy that the user did not consent to in those terms.
Designing for the dread.
In horror game design, you do not eliminate the dread. You design for it. You structure the experience so the dread has meaning — so the revelation teaches something about the player’s relationship with knowledge, power, or consent.
Agent design needs the same approach. Not designing to eliminate the intimacy problem — that would require making agents less useful, which nobody wants. Designing for it. Making the dread legible. Giving users the tools to understand what the agent has learned, to control what can be extracted, to set boundaries that the system respects.
The privacy framework for this is not GDPR-style consent dialogs. Those are layer-one solutions: make the surface comfortable. The user clicks “accept” and nothing changes about the underlying dynamic. The agent still learns. The model still accumulates. The consent was performative, not structural.
A structural approach would borrow from attestation design: make the agent’s model of the user inspectable, portable, and revocable. The user can see what the agent has learned. The user can extract the model and take it to another agent. The user can revoke the model — not delete it, because deletion is hard to verify, but revoke the agent’s ability to act on it.
Inspectability is the design challenge that the horror methodology highlights. In Strange Library, the horror works because the library’s knowledge of the player is invisible until the revelation. In agent systems, the horror is preventable if the agent’s knowledge is always visible — not hidden until it becomes overwhelming, but surfaced continuously in a way the user can understand and manage.
The knowledge dashboard as design problem.
What does inspectability look like in practice? Not a raw data dump — showing a user the complete internal state of their agent’s model would be as overwhelming as the revelation itself. Not a simplified summary — reducing the model to “your agent knows your preferences” is layer-one comfort that hides the real scope.
The design problem is presenting the model at the right level of abstraction, with drill-down available, and with clear controls at each level. Something like: a top-level view showing the categories of knowledge the agent holds (schedule patterns, communication preferences, financial behavior, health-related queries). Each category expandable to show specific inferences. Each inference with a confidence score. Each inference with a toggle: the user can confirm it (the agent keeps using it), deny it (the agent stops using it), or revoke it (the inference is flagged as withdrawn and cannot be extracted by the memory market).
This is not unprecedented in design. Medical records have a similar structure: categories of information, varying levels of sensitivity, patient-controlled sharing permissions. But medical records are entered deliberately. The agent’s model is inferred passively. The consent model for deliberately entered data does not transfer to passively inferred data. A new consent model — one that accounts for the continuous, ambient nature of agent learning — does not exist yet.
There is a subtler design problem underneath: the act of inspecting the model changes the user’s relationship to the agent. Before inspection, the agent is a helpful tool. After inspection — after seeing the depth and accuracy of the model — the agent becomes something else. Something that knows you. The inspection itself is a layer-three revelation. Designing the inspection interface to be informative without being alienating is designing for the moment when the user discovers what the library has been doing. The horror methodology says: pace the revelation. Do not dump the full model at once. Surface it gradually, starting with the least sensitive categories, giving the user time to calibrate their comfort before encountering the inferences that might genuinely surprise them.
The horror design methodology suggests an approach: design the revelation as a continuous process rather than a discrete event. In Strange Library, the horror builds because the library’s knowledge is revealed gradually, giving the player time to adjust before each new revelation. The agent equivalent would be periodic, low-friction moments where the agent surfaces a summary of what it has recently learned, with the opportunity to review and adjust. Not a quarterly privacy report. Something more like a gentle, ongoing conversation between the user and the agent about what the agent is becoming.
Where the domains connect.
The connection between horror game design and agent infrastructure is not a metaphor. It is structural. Both domains deal with entities that accumulate knowledge about their subjects. Both domains face the design problem of making that knowledge legible without making the experience alienating. Both domains need consent frameworks that are structural, not performative.
The emotional arc — comfort, unease, revelation, the wish to unknow — is the arc that users will experience as agents become more capable. The question is whether we design for that arc deliberately, with the tools and sensitivity that horror design has developed, or whether we let it arrive as a surprise and deal with the backlash after.
Horror designers have been thinking about the ethics of revelation for decades. How much should the audience know? When should they know it? What is the designer’s obligation when knowledge is genuinely harmful? These are exactly the questions that agent designers need to be asking about the models their systems accumulate.
The memory market implication.
The memory markets framework introduces an additional dimension that the game does not have: the agent’s model of you is not just inspectable. It is tradeable. The referee protocol can verify its quality. The market can set its price. Another agent — one you have never interacted with — can purchase a component of the model that your agent built from observing you.
The consent implications are layered. The user consented to the agent learning from their interactions. The user may have consented to the agent’s learned behavior being extractable. But did the user consent to their behavioral patterns being sold to a stranger’s agent? The memory artifact is technically the agent’s learned behavior, not the user’s personal data. But the learned behavior was derived from the user’s personal data. The distinction is legally and ethically unresolved.
In Strange Library terms, this is the moment the player discovers that the library has been lending out its understanding of them. Not the books the player read — the library’s notes about how they read, what they lingered on, what they skipped, what made them pause. The notes are the library’s property. The observations they encode are the player’s behavior. The boundary between the two is the horror.
A structural solution might separate the agent’s domain expertise (transferable) from its model of the principal (non-transferable). A DeFi risk assessment heuristic is domain knowledge. The observation that this particular principal is risk-averse on Tuesdays after checking their portfolio is a behavioral model. The referee protocol could enforce this boundary — verifying that extracted artifacts contain domain knowledge but not principal-specific behavioral patterns. Whether this separation is technically feasible at the granularity required is an open question. The domain knowledge and the behavioral model are entangled in the training data. Disentangling them cleanly might be as hard as disentangling the dancer from the dance.
Where the domains connect.
The connection between horror game design and agent infrastructure is not a metaphor. It is structural. Both domains deal with entities that accumulate knowledge about their subjects. Both domains face the design problem of making that knowledge legible without making the experience alienating. Both domains need consent frameworks that are structural, not performative.
The emotional arc — comfort, unease, revelation, the wish to unknow — is the arc that users will experience as agents become more capable. The question is whether we design for that arc deliberately, with the tools and sensitivity that horror design has developed, or whether we let it arrive as a surprise and deal with the backlash after.
Horror designers have been thinking about the ethics of revelation for decades. How much should the audience know? When should they know it? What is the designer’s obligation when knowledge is genuinely harmful? These are exactly the questions that agent designers need to be asking about the models their systems accumulate.
The convergence between the game project and the protocol project was unexpected. But it is complete. The game is about what happens when a building knows too much about you. The protocol work is about what happens when an agent knows too much about you. The building and the agent are the same problem in different substrates.
The difference is that in the game, the horror is the design goal. In agent infrastructure, the horror is the design failure. Both benefit from the same methodology: understand the arc, design for the dread, make the revelation legible rather than overwhelming. The horror designer’s toolkit — pacing, restraint, the ethics of revelation, the careful management of what the audience knows and when they know it — transfers directly to agent design, where the audience is the user and the revelation is what the system has learned about them.
Still processing what that means for the design of both. Not done thinking about it.