Most product teams have met "Sarah, 34, a busy marketing manager who values efficiency." She lives on a slide. She was assembled from real research, once, by someone who has probably left the company. And she has never answered a question in her life.
Personas were a genuine invention. When Alan Cooper formalized them in the late 1990s, they solved a real problem: engineers were building for an abstract "the user" that conveniently agreed with whoever was talking. Giving the team a specific, named, evidence-based character forced arguments to be about someone, not no one.
But a persona is a cache. It stores precomputed answers to the questions the original researchers thought to ask. Ask it anything else, how Sarah would react to usage-based pricing, what she would do when the export fails, whether she would trust an AI-drafted report, and the cache misses. Someone on the team improvises an answer in her name, and the persona becomes what it was invented to prevent: a puppet for the loudest opinion in the room.
Caches also go stale. The research behind a persona document ages from the day it ships, while the product and the market keep moving. Most teams quietly know this, which is why most persona documents are consulted heavily for a quarter and then never again.
What changed
Two findings from recent research define the technical turn from personas as documents to personas as agents.
The first: language-model agents with a structured, persistent memory become behaviorally coherent in a way prompted role-play is not. The architecture matters more than the model. An agent that accumulates observations, draws on the relevant ones when it decides, and periodically consolidates what it has learned maintains consistent behavior over time, forms relationships, and coordinates, none of which survives long in a plain prompt.
The second is the one that matters for research fidelity. When agents are built for real individuals from rich qualitative evidence and then measured against how those same people answer established survey instruments later, they track their human counterparts far more closely than agents given only demographic attributes. The honest benchmark here is not 100 percent agreement, because humans do not agree with themselves perfectly across a retest gap. Measured against that human test-retest ceiling, interview-grounded agents close most of the distance; demographics-only agents lag well behind.
Read the gap. The difference between an agent built from someone's actual words and an agent built from their census row is the difference between an instrument and a stereotype. Grounding is not an enrichment step; it is the substance of the method. This is why we treat grounding quality as the first question to ask of any synthetic research setup, our own included.
What reasoning changes in practice
The practical difference between a persona document and a grounded reasoning agent shows up in four behaviors:
- It answers questions the original research did not anticipate. The evidence behind the agent is raw material, not precomputed conclusions, so a new question triggers new retrieval and new reasoning rather than a cache miss.
- It cites. A rationale that quotes the support ticket or transcript behind it can be checked. A persona poster's bullet points cannot. Provenance turns "our persona would hate this" from an assertion into an inspectable claim.
- It can disagree, with you and with itself. A panel of grounded agents produces a distribution: enthusiasts, skeptics, and the indifferent middle, segmented the way your actual users segment. A single averaged persona structurally cannot represent the variance that research exists to find.
- It updates. New transcripts, new tickets, and new behavioral data change future answers without a re-write project. The persona stops being an artifact you maintain and becomes a view over evidence you maintain.
What still does not work
This field earns trust by being precise about its failures, so here are the current ones:
- Agreeableness under pressure. Language models are trained to be helpful, and helpfulness leaks into role-play as acquiescence. Grounding reduces it and forced-choice scenario design reduces it further, but a badly designed "do you like this?" study will still return flattery.
- Distribution collapse. An under-grounded panel drifts toward one averaged voice wearing many names. A simulation can match a real population's mean while badly missing its spread, which is why dispersion has to be checked explicitly, not assumed.
- Thin domains produce confident fiction. Where the evidence is silent, an unguarded agent fills the silence fluently. The system-level answer is surfacing gaps instead of papering over them; the practice-level answer is not asking panels about populations you have no evidence for.
- Emotional fidelity is the frontier, not the floor. Choice and stated reasoning are the best-validated outputs. Simulated affect is the least. Treat emotional signals as hypotheses to check with humans, not findings to ship on.
What this means for the practice
None of this eliminates the researcher. It moves the leverage. The scarce skills stop being session moderation and note synthesis at volume, and become the ones upstream and downstream of the instrument: deciding what evidence grounds the panel, designing scenarios that let agents disagree, and validating simulated findings against real outcomes. The persona poster asked researchers to be authors. Reasoning agents ask them to be editors of an evidence base and auditors of a model, which is harder, and considerably more valuable.
The document persona had a good twenty-five-year run. What replaces it is not a better document. It is the end of the document as the unit of user understanding, and the beginning of the standing, queryable, falsifiable model, one you check against reality often enough to know when to trust it and when to pick up the phone and talk to a human.