THEME 01PersonaOS

The Agentic AI Era

Agents that decide like the people they act for

Agents now plan, prioritize and decline on someone's behalf. PersonaOS encodes a person's decision criteria in seven layers—and records where those criteria conflict and when the person overrides them—so the agent chooses what that person would have chosen.

Why tone-based personas break down in agents

  • 01

    Written for conversation

    Their target is how a person talks, not how the person ranks options. They say little about the decisions an agent must make.

  • 02

    No structure

    Temperament, values, habits and wording sit in one undifferentiated text. When two statements pull in different directions, nothing says which one governs the decision at hand.

  • 03

    Contradiction removed

    Internal conflict is left out to keep the character consistent. Real people hold competing values and act against their preferences when a situation demands it. An agent that follows tendencies alone gets easy decisions right and hard ones wrong.

Seven layers of everyday judgment

Lower layers change slowly and sit close to biology; upper layers change quickly and show up in behavior. Select a layer to see what it records and the research behind it.

↑ Fast-changing · observable

↓ Slow-changing · biological basis

↕ Vertical dimension: contradictions and situational rules run through all seven layers

L7 · Expression

Expression

Vocabulary, sentence style, register, how emotion is expressed

Grounding
Language as an individual difference (Pennebaker & King, 1999)
L6 · Narrative

Narrative

How the person explains their life, defining events, and where they are headed

Grounding
McAdams' life-story model; identity development (Erikson, Marcia)
L5 · Relations

Relations

Relationships, group membership, what is shared with whom (disclosure scope)

Grounding
Social identity (Tajfel & Turner); identity theory (Burke & Stets)
L4 · Competence

Competence

Domain knowledge, skills, working methods

Grounding
Deliberate practice (Ericsson); skill stages (Dreyfus)
L3 · Operation

Operation

Habits, routines, decision styles

Grounding
Habit research (Wood & Neal); dual-process theory (Kahneman)
L2 · Character

Character

Values, beliefs, ranked motives

Grounding
Character dimensions (Cloninger); characteristic adaptations (McAdams)
L1 · Core

Core

Temperament, cognitive style, physiological rhythm

Grounding
Temperament model (Cloninger); five-factor model; attachment (Bowlby)

Layers consulted by question type

The correction list and L7 documents are always included

  • Weight sits in L2 and L3

    Most of a decision rests on what the person wants (L2) and how they usually choose (L3). L7 is where a decision becomes visible, not where it is made.

  • Layers inform each other

    Lower layers bound higher ones: a temperament high in novelty seeking (L1) makes quick, exploratory habits (L3) likely. Higher layers are evidence about lower ones: a phrase repeated in L7 is a clue to a value in L2.

  • Disclosure scope in L5

    Each relationship carries a disclosure scope, so an agent acting in front of several parties withholds what the person would withhold. Without it, a faithful digital twin becomes a leak.

The vertical dimension: where judgment gets hard

The seven layers capture a person's defaults. Hard decisions happen when those defaults collide, or when a situation calls for acting out of character. The vertical dimension runs through all seven layers to handle exactly these moments, in two parts.

Contradictions

Map the contradictions

A contradiction can sit inside one layer or span two, so it is recorded on its own and linked to every element involved. We keep contradictions on record because they mark exactly where an agent that follows typical tendencies is most likely to get it wrong.

  • Internal

    Within a layer

    Two elements of one layer conflict

    Values both candour and harmony, so frank feedback and keeping the peace compete (both L2)

  • Crossing

    Across layers

    Elements of different layers conflict

    Health is the stated first priority (L2), but sleep is routinely cut to finish work (L3)

  • Shadow

    Unacknowledged

    An aspect the person does not acknowledge

    A need for approval in someone who describes themselves as indifferent to others' opinions

  • Unresolved

    Open question

    A question the person has not answered

    Whether to stay in a stable job or start a company

Grounding: cognitive dissonance (Festinger), the shadow (Jung), the dialogical self (Hermans), identity crisis (Erikson), self-discrepancy theory (Higgins)

Situational rules

Codify the moments people act against type

Sometimes people act against their own temperament: they take on work they'd rather refuse because the deadline is theirs, speak up when they'd rather stay quiet, or put someone else's safety first. A persona that misses these moments misses the decisions that matter most.

  1. 01ConditionWhen the rule appliese.g.A client deliverable she is accountable for is at risk
  2. 02OverridesThe default criterion it sets asidee.g.Declining collaboration requests
  3. 03ActionWhat the person does insteade.g.Accept and join the team until the deliverable is safe
  4. 04ReturnWhen the default comes backe.g.Once the deliverable ships

When two rules hold at once, the higher-priority rule decides. If priorities are equal or missing, the agent reports the conflict instead of choosing.

Grounding: the cognitive–affective personality system (Mischel & Shoda: consistency lies in if–then patterns), traits as distributions of states (Fleeson), free-trait theory (Little: an introvert acts as an extravert for projects that matter)

How a PersonaOS agent reaches a decision

The agent forms a first judgment from the person's defaults, then checks two things: does it touch a recorded contradiction, and does a situational rule apply? If a rule applies, the rule decides, and the agent explains that it is departing from the usual preference and why.

Example: a fictional project lead

L1 · L3
Strong preference for working alone; habit of declining unplanned meetings
L2
Keeping commitments to clients is her highest work value
Crossing contradiction
Solitary working style ↔ client commitment
Situational rule
If a client deliverable she owns is at risk, accept collaboration; return once it ships

A fictional person, used only to illustrate the format.

Pick a situation

PersonaOS agent

I'll pass on this one. I need to stay on my own work.

The rule's condition (client deliverable at risk) is not met, so the default judgment stands.

PersonaOS agent

I would normally decline this, but the deadline is my responsibility, so I will take it this time.

The rule's condition is met, so the rule decides. The agent states that it departs from her usual preference and why.

Agent with only a profile that says "prefers to work alone"Expected to decline both times.
Request in context
Retrieve layers for this question type
First judgment from default criteria (L1–L7)
Touches a recorded contradiction?
yes ↓ · no → Act on the default judgment
Situational rule condition met?
yes
Act on the rule; state the departure
no
Act on the default judgment
Log → review → correct

↺ Corrections update the documents that later requests retrieve

Define. Document. Deploy and verify.

Every mismatch becomes a correction, and every correction updates the documents.

  1. 01

    Define

    Interviews, the person's own writing, standardized test results and records of past decisions are broken down into the seven layers and the vertical dimension.

  2. 02

    Document

    One element per machine-readable document, with layer, type, source grade, confidence and links. Earlier versions are kept with the date and reason of each change.

  3. 03

    Inject and verify

    Only the documents relevant to a request go into the agent's context. The person or a proxy reviews the result.

↺ Mismatch → correction → document updated

Source grades and correction first

Every statement is graded by its source. A direct correction by the person overrides every other source.

  • When two graded statements conflict and no correction resolves them, the conflict is not settled silently; it is recorded as a contradiction.
  • Inferred statements are never promoted to a higher grade because a model repeated them.
  • Purpose: to prevent generalizing a whole personality from a single document, and attributing a plausible deeper motive the person does not recognize.
  1. CorrectionChanged by the person. Overrides every grade and is never deleted without the person's instruction.First
  2. DirectStated by the person
  3. MeasuredRecorded by a standardized instrument or record
  4. ObservedSeen repeatedly in behaviour
  5. InferredDerived by a researcher or a model

One framework, three kinds of persona

What changes is how much of the structure each persona carries.

  • 01

    Complete Digital Twin

    One person, all seven layers and the vertical dimension

    • L1
    • L2
    • L3
    • L4
    • L5
    • L6
    • L7
    • Con
    • Rule

    Judges as that person would—for decision support or as a record of a person.

  • 02

    Purposeful Persona Extract

    Only what a task needs

    • L1
    • L2
    • L3
    • L4
    • L5
    • L6
    • L7
    • Con
    • Rule

    Selected by period of life, role, strengths or situation. A founder in the first year and the same founder a decade later decide differently; an extract can choose which to reproduce.

  • 03

    Synthetic Ideal Agent

    Elements of several people, or of none

    • L1
    • L2
    • L3
    • L4
    • L5
    • L6
    • L7
    • Con
    • Rule

    A persona designed for a role. Elements that do not fit together are recorded as contradictions, so the coherence of the synthesis can be inspected.

Always includedSelected for the purpose

Five predictions we will test

Not yet tested · no results

We build four conditions for the same person on the same model and score each against what the person actually decided. The bars show the predicted order only; no results have been reported yet.

  • C0No persona
  • C1~10-sentence profile
  • C2Seven layers, no vertical dimension
  • C3Full PersonaOS

An ablation of C3 without situational rules separates their effect from that of contradictions.

  • H1Agreement with the person's past decisions increases from C0 to C3
  • H2The advantage of C3 over C2 is concentrated in decisions with a conflict between values, and small elsewhere
  • H3The correction-override rule raises the person's own rating of the agent's answers
  • H4Recording disclosure scope in L5 reduces how often the agent reveals what the person would withhold
  • H5Situational rules raise agreement on decisions where the person acted against their disposition; without them, agreement falls on exactly these

Measures

  • Decision agreement
  • Blind identification
  • Self-rating (1–5)
  • Profile distance
  • Disclosure adherence

Open questions

  1. 01This is a framework and a protocol. No experimental results are reported, and none of the predictions has been tested.
  2. 02It was developed on a small number of cases; layer boundaries need testing across cultures, occupations and ages.
  3. 03Some elements sit on a boundary—a recurring decision principle can read as a value (L2) or a habit (L3)—and clearer allocation rules are needed.
  4. 04Documentation is costly. Wide use will need automated interviewing and drafting.
  5. 05Self-reports reflect the self a person wishes to present. Measured and observed grades reduce this bias but do not remove it.
  6. 06There is not yet a principled way to set priorities between rules whose conditions hold at once.

Ethical commitments

  • 01

    Consent

    No twin without explicit consent. The same applies to real people used in synthetic personas.

  • 02

    Control

    The person can read, correct and delete their documents. The correction-override rule builds this into the design.

  • 03

    Disclosure

    L5 disclosure scopes are enforced, not advisory.

  • 04

    Labelling

    People interacting with a twin can tell that it is not the person.

  • 05

    Sensitive data

    Health, financial and family information gets a separate level of protection.

The test of a persona for an agent is not whether it sounds like someone. It is whether it chooses what that person would have chosen, including the times they chose against themselves.

Personas for agentic AI must encode judgement, not tone · ddai, 2026 · A framework and evaluation protocol; no experimental results yet.

Research behind this design

  • HarvardAgents

    Agents repeat their past experiences, mistakes included

    Harvard · Michigan State · University of Minnesota · Zidi Xiong, Yuping Lin, Wenya Xie et al. · ACL 2026

    The study tests how adding and deleting memories shapes four LLM agents over time. Agents follow similar past experiences, so stored errors spread; keeping only runs that matched the ground truth raised EHRAgent accuracy from 13.05 (storing everything) to 38.50.

    ddai insight

    For an agent, memory works less like a record and more like an example to imitate. Choosing what to keep is choosing how it will judge next.

    How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior

    Read the paper
    PersonaOS
  • KAISTAgents

    In role conflicts, LLMs follow role preferences over context

    KAIST · Jisu Shin, Alice Oh et al. · Findings of ACL 2026

    The team built 13,914 role-conflict scenarios and used situational urgency as an objective baseline for LLM judgment. The 10 models tested deviated substantially; their choices rested mainly on preferences for particular social roles.

    ddai insight

    An AI's judgment already carries a ranking of roles. That ranking can outweigh what the situation calls for, so designing judgment criteria starts with bringing it to the surface.

    RoleConflictBench: A Benchmark of Role Conflict Scenarios for Evaluating LLMs' Contextual Sensitivity

    Read the paper
    PersonaOS
  • StanfordAgents

    Agents built from self-reports predict each person's answers

    Stanford University · Joon Sung Park, Michael S. Bernstein et al. · 2026

    The team built an LLM agent for each of 1,052 Americans from that person's interview and survey answers. On held-out GSS items, agents reached 83% (interviews), 82% (surveys), and 86% (both) of each person's two-week retest consistency, versus 74% for demographics alone.

    ddai insight

    An AI modeled on one person needs that person's own answers, not just the groups they belong to. Demographics-only agents showed wider accuracy gaps across racial and ideological groups.

    LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals

    Read the paper
    PersonaOS
See all research

Work

The expertise inside the work

  • 01

    AI assistant persona, scaled to 41 languages

    Expertise
    Persona design, response style, and per-language rewriting
    Method
    Spec first; native writers rewrite to the same spec; one rubric scores every language
    Deliverables
    Persona spec, multilingual responses, and evaluation criteria
    PersonaOSLocalization
  • 03

    Judgment criteria for AI agents

    Expertise
    PersonaOS-style judgment design
    Method
    Record a person's criteria layer by layer, then write rules for where they conflict
    Deliverables
    Judgment criteria document and evaluation set
    PersonaOS
See all work
Next: 02Humanoid
Talk to us about PersonaOS
  • Contact us

    Building a robot or an agent? ddai designs its character and judgment.
    From the persona spec to evaluation data in 41 languages, one team sees it through.

    Start a project
  • ddai

    Before AI can stand beside people, it needs a character and a way to judge. ddai designs both.

    Meet ddai