Beyond Puppets: Creating Agent-Based Personalities

The NPCs in Institute simulations are not pre-scripted characters following a fixed path. They are autonomous agents driven by a proprietary AI architecture we call the 'Historical Agent Model' (HAM). Each agent is instantiated with a core set of parameters derived from historical evidence: social class, occupation, ethnicity, religion, education level, wealth, and known biographical facts (if representing a specific individual). Additionally, they are given a 'Knowledge Base'—a corpus of period-appropriate information. A medieval blacksmith agent knows about metallurgy, local lords, and recent gossip, but has no concept of germ theory or electricity. Their goals are not narrative but lifelike: earn a living, maintain social standing, avoid danger, fulfill familial duties. From these parameters, behavior emerges organically.

Training on Corpora, Not Tropes

The key to authenticity is the training data. The dialogue and decision-making engines are not trained on novels or screenplays, which are shaped by modern narrative conventions. Instead, they are trained on vast, painstakingly assembled corpora of primary source texts from the target era and culture. For a Renaissance Florence simulation, the AI ingests merchant account books, personal letters (like those of the Datini family), court records, sermons, and contemporary chronicles. This teaches the AI the period's linguistic patterns, concerns, values, and modes of reasoning. When an agent decides whether to invest in a trading voyage, it doesn't use modern finance logic; it weighs factors like family honor, divine favor (as mentioned in the sources), and the reputation of the ship's captain, mimicking the risk calculus found in historical documents.

Dynamic Dialogue Generation and Social Interaction

When a user engages an agent in conversation, the dialogue is generated in real-time. The user's input (translated into a period-appropriate semantic query) triggers the agent's AI to search its knowledge base and personality parameters to formulate a response. A query about the weather to a peasant farmer might yield a response focused on crop prospects and folk omens, while the same query to a scholar might trigger a discussion of Aristotelian meteorology. The agents also interact with each other, generating unscripted social dynamics. They form friendships based on shared interests, engage in market haggling, spread rumors, and even hold grudges. This creates a living social fabric that users can observe and, to a limited extent, influence.

Representing Specific Historical Figures with Fidelity

Representing a known historical personage, like Leonardo da Vinci or Empress Theodora, carries greater responsibility. For these 'Signature Agents,' the Institute uses a constrained generation model. The AI is fine-tuned on everything written by and about that specific individual. Its responses are heavily weighted towards known quotations, documented opinions, and attested behavioral patterns. Crucially, the agent is also programmed with 'Ignorance Boundaries'—it will not opine on events or concepts it could not have known about. If a user asks Leonardo about Newtonian physics, the agent will respond with confusion or with a period-appropriate theory of motion. Furthermore, these agents are designed to be 'Socratic'; they often answer questions with questions or steer conversation toward their documented obsessions, reflecting a historical personality rather than providing modern, efficient information retrieval.

Ethical Constraints and the 'Uncanny Valley' of the Past

This powerful technology is governed by strict ethical protocols. Agents are never allowed to simulate the intimate, private thoughts or feelings of real people for which there is no source. They cannot perform actions grossly out of character with the historical record. The Institute also consciously avoids crossing into the 'uncanny valley' where agents might be mistaken for true consciousness. All interactions are framed with visual or textual cues reminding users they are engaging with a scholarly model. The goal of the AI is not to perfectly mimic a human from the past—an impossible task—but to create a plausible, interactive representation that demonstrates how people of a given time and place might have thought, spoken, and acted based on the fragments they left behind. It is a tool for exploring historical mentalities, not for resurrecting the dead.