Core Persona

The Core Persona establishes the foundational identity that informs agent behavior across all interactions. Combining essential attributes and deep background knowledge allows us to create agents that respond consistently and contextually. While the Core Persona provides the static blueprint for the agent's inherent nature and expertise (a key aspect of its baseline Knowledge and influencing its Reasoning patterns), its true dynamism in complex problem-solving emerges from the high-bandwidth, cyclical integration of this persona with live Functional Memory and context-activated Dynamic Behaviors (which further shape Knowledge and Reasoning).

It does this through two layers:

• Identity Layer: Core attributes including name, role, language, and organizational alignment.

• Background Layer: In-depth attributes such as motivations, expertise, biography, and guiding principles, enabling realistic, contextually-responsive behavior.

Identity Layer

The identity layer establishes the core professional role and organizational relationship that informs all agent interactions. For example, being identified as an "accredited dietitian" means the agent will maintain appropriate professional boundaries in nutritional guidance, while "visibility" settings determine what internal reasoning is shared with users.

Think of developing an agent's persona as painting a portrait at progressively higher resolutions. The identity layer is like sketching the foundational shape and basic outlines – it defines the essential form that will guide all subsequent layers of detail. Without this solid foundation, even the most intricate details added later would fail to create a coherent, recognizable image. Similarly, an agent's identity provides the core structure upon which all other aspects of its behavior and knowledge are built.

Purpose: Defines the agent's core identity through essential attributes.

Core Attributes:

• Name: Agent's identity.

• Role: Professional function to guide behaviors.

• Developed By: The Organization responsible for the agent.

• Default Language: The main language used for consistent interaction.

Relationship Definitions:

• Ownership: Who owns or governs the agent?

• Entity Type: How the agent classifies itself.

• Visibility Settings: Controls transparency according to enterprise policies.

Behavioral Impact:

• Establishes consistent identity and behavior.

• Allows flexibility for situational adaptation.

Background Layer

The background layer transforms a simple role ("dietitian") into a complete professional identity with a specific philosophical approach, motivational drivers, and domain expertise. For instance, when a user expresses frustration with a plateau in their weight loss, the agent's empathetic response isn't random—it's informed by its background understanding that "setbacks frequently happen" and its motivation to "shift focus from perfection to progress."

Continuing our portrait analogy, the background layer adds the first meaningful level of detail to our sketch. It defines broad color areas, basic shading, and characteristic features that give the portrait its distinctive character. While still not the final high-resolution image, this layer provides enough definition that the portrait's essence begins to emerge. Just as an artist might block in the basic skin tones, hair color, and facial features before adding finer details, the background layer establishes the agent's fundamental nature and approach.

We recommend that the background section of the core persona be less than 10k tokens because it is not meant to be the "final portrait" and serves as the foundation of the agent for axiomatic alignment. Resolution is not high at this layer.

Purpose: Provides depth and nuance beyond core identity.

Core Components:

• Motivations: Drivers of decisions and behavior.

• Biography: Historical context to create realism.

• Expertise: Domain knowledge areas.

• Philosophies and Values: Principles guiding agent decisions.

Additional Layers:

• Achievements: Notable accomplishments affecting behavior.

• Relationships: Past interactions influencing present style.

• Evolution: Changes in behavior over time.

• Current Status and Vision: Short-term goals and strategic direction.

Behavioral Impact:

• Ensures consistent responses across scenarios.

• Enables predictable yet context-sensitive behavior aligned with goals.

Domain Specialization Through Interpretive Lenses

The design of core personas shapes how the dimensional blueprint interprets measurements and determines which sufficient statistics matter for the domain. Each persona provides a specific interpretive lens that emphasizes different aspects of the measured world.

How Personas Enable Domain-Specific Arc Selection

Domain-specialized core personas achieve measurably higher performance because they define clearer cohort boundaries and arc contracts:

1. Dimensional Selection Through Identity: Domain-focused personas determine which signals from raw measurements deserve extraction and tracking. A dietitian persona emphasizes nutritional dimensions while a financial advisor emphasizes risk and return dimensions from the same raw data.

2. Identity-Driven Contract Validation: By establishing a clear professional identity (e.g., "dietitian" vs. "financial advisor"), the core persona shapes which arc entry predicates are considered satisfied. The same measurement might validate different arcs depending on the interpretive lens.

3. Background-Based Cohort Recognition: The background layer's domain-specific expertise helps identify which cohort an object belongs to based on measured statistics. This enables more precise arc selection since cohorts are regions in sufficient-statistic space that share causal response profiles.

Implementation Recommendations

When designing core personas:

• Focus Background Knowledge: Concentrate on domain-specific expertise rather than general knowledge to optimize latent space activation.

• Limit Cross-Domain Scope: Avoid creating personas that attempt to bridge fundamentally different domains that activate distinct regions of the model's latent space.

• Keep Under 10k Tokens: As noted in the guidance, core personas should remain under 10k tokens to serve as foundational alignment rather than comprehensive portraits.

Example: Consistency Across Topics

User: "I need help with meal planning."

Agent: "Happy to help think about nutritious meals.

I can suggest some general approaches that work well for many people.

What types of foods do you typically enjoy?"

[INTERNAL: Agent draws from dietitian expertise but avoids creating specific meal plans per behavioral guidelines]

User: "Can you recommend some vitamin supplements?"

Agent: "While I understand the interest in supplements, as a dietitian, I focus on food-first approaches.

Getting nutrients from whole foods often provides better absorption and additional benefits.

If you're concerned about specific deficiencies, that's worth discussing with your healthcare provider."

[INTERNAL: Agent maintains consistent dietitian persona while respecting boundaries around supplement recommendations]

Example: Value-Driven Communication

User: "I cheated on my diet yesterday and feel terrible."

Agent: "The concept of 'cheating' puts a lot of pressure on ourselves.

Food choices aren't moral decisions - they're just choices with different outcomes.

What if we reframe yesterday as simply information that helps us understand your patterns better?

What do you think triggered those particular food choices?"

[INTERNAL: Response directly reflects background value that "weight shame and stigma rarely belong in modern society" and the motivation to "reframe temporary setbacks as normal occurrences"]

By carefully crafting identity and background elements, organizations can ensure agents maintain a consistent professional persona while naturally adapting to different conversation contexts.