The individual
How one person works
Their workflows, the tools they reach for, what they adopted, kept and dropped, and where AI already helps. Learned from traces of real work they choose to share, not from surveys.
Research
A world model of how people work with AI.
We study computational and latent representations of individuals and teams: how they work, what they know, and how they use AI. Inspired by world models and latent reasoning, built from real work, and open to the people they describe.
The question
AI systems now learn models of the world compact enough to imagine with: they predict what happens next before they act. Organisations have nothing like that for themselves. What a person does well, which tools they rely on, what they tried and dropped, and how a practice travels from one colleague to the next live in habits and memory, and they disappear when people move on.
We are asking what the world model of a person at work would be, and of a team. Can a learned representation say which next step will help someone, and test it in simulation before anyone spends a week on it?
Three things we represent
The team
How knowledge moves
Who knows what, which practices spread and which stall, and who a newcomer should learn from. Teams carry knowledge between people; we want a representation of that, not a list of individuals.
The landscape
What AI can do now
Tools, models and ways of working, connected by what builds on what, and changing every week. The map on our front page is the first public view of it.
Why latent
Written records capture what can be written down: checklists, documents, job descriptions. Most of how people actually work cannot be: the order they do things in, when they trust an answer, what they reach for under pressure. World models learn a compact hidden state that predicts what comes next, and latent reasoning lets a model think in that state instead of in words. We are testing whether the same ideas give a representation of a person or a team that is small, predictive, and still explainable.
What we are testing
- Prediction. Does a learned representation predict which new tools a person will still be using a month after trying them, better than rules or self-reports do?
- Simulation. Can a team’s representation be rolled forward to estimate how a practice would spread, and does that hold up against real pilots?
- Transfer. When a new model or tool appears, can the representation say whom it helps before anyone tries it?
- Legibility. Can every prediction be traced back to evidence a person can read, question and correct?
Principles
A representation of a person belongs to that person. It is built only from what they choose to share, kept on their machine by default, and they can see it, correct it and delete it. It exists to suggest what helps next. It is never used to rate, rank or watch people.
Where it stands
Today Harklane runs on explicit structure: a catalog of tools and ways of working, the prerequisites between them, and evidence from a person’s own sessions. The latent representations are research. They are early, and nothing on this page is a product claim.
If you work on world models, representation learning or how organisations learn, we would like to hear from you.
Work we read
Outside work that shapes our questions. These are other people’s findings, not our results.
Ha and Schmidhuber · March 2018
World Models
An agent learns a compressed model of its environment and can train inside its own simulation of it. The starting point for our question: what is the world model of a person at work?
Hafner et al. · January 2023
Mastering Diverse Domains through World Models
DreamerV3 plans by imagining futures in a learned latent space, with one setup across many domains. We borrow the idea of rolling a state forward before acting.
Assran et al. · January 2023
Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture
I-JEPA predicts in representation space rather than raw detail, so the model keeps what matters and drops the rest. A useful stance for modelling people, where most detail is noise.
Hao et al. · December 2024
Training Large Language Models to Reason in a Continuous Latent Space
Coconut lets a model reason in its own continuous state instead of in words. It suggests reasoning about a person’s work need not pass through written summaries at every step.
Zou et al. · November 2025
Latent Collaboration in Multi-Agent Systems
Agents share internal representations instead of text. An early sign of what a team-level representation could look like, though not yet one that works across any model.
Park et al. · April 2023
Generative Agents: Interactive Simulacra of Human Behavior
Agents with memory, reflection and planning produce believable individual and social behaviour. Shows what a representation of a person can support, and what it cannot yet be trusted with.
Park et al. · November 2024
LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals
Agents built from interviews with over a thousand real people reproduce their answers closely. Evidence that individuals can be represented well enough to simulate, with consent.
Xu et al. · December 2024
TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks
A simulated company for testing agents on connected workplace tasks. A reminder that work is a system, and that failures appear between tasks, not inside them.