The AR Labs vision

AR Labs is working toward a way of using AI in which a person speaks naturally and the system answers with structure they can see: maps of ideas, goals and work rather than walls of text. Underneath that interface sits a cognitive architecture, a way of organising memory, context and attention that plays to the strengths of both people and language models and works around their weaknesses. We want this capability in the hands of individuals and small teams, not only of large organisations.

Those three parts, the interface, the architecture and who holds it, make up our vision. This page sets out each, then the longer horizons we see beyond them.

Voice in, concepts out

We start from a view: that conversation will become a main way people work with computers, and that devices and applications will be controlled more and more through voice and dialogue. So we treat voice as the primary input, and aim for an AI that understands what a person means, not only the words they used.

Output is the other half. A spoken or written answer is linear, and a long one overloads the small working memory a person thinks with. So our intended primary output is conceptual visualisation: mind maps, pictures of how concepts connect, goal trees and structured summaries that a person can explore and return to. As we put it, conversation handles nuance and intent, and visualisation handles structure and complexity. The page on VOX, the lab's working system, describes this interface in detail, including a design requirement, not yet built, that a branching conversation be shown as a goal tree the person can fork, step back up and continue from.

Our live voice work so far is Salesbot, a browser voice agent. The full combination, a conversation drawn as concepts while it happens, is the goal rather than a built product.

A cognitive architecture for two kinds of mind

We describe what we build as cognitive AI systems: AI agents (language models run in a loop with tools, so they can act, not only reply) with rich memory and context retention, coherent reasoning and planning, an understanding of what users mean, and visual feedback that shows how the agents think and work.

What makes those systems usable, we think, is the architecture around the model rather than the model alone. We look for three properties in that architecture:

• Structured context that focuses the model's attention. What a model is given to read shapes what it attends to. Our discipline of context engineering, choosing what goes into a model's context window at each step, follows from this.
• Information held at several levels, traversable through conversation. A person or an agent starts from the general picture and descends to detail only when the work needs it, the same progressive disclosure we use to organise our own documents.
• Graphical interfaces that respect human cognitive limits. A few things at a time, with the rest held in a picture the person can return to.

The architecture draws on our founding framework, cognitive ecology, which treats both a person and a model as minds with a small span of attention in front of a large associative store.

Capability for individuals

The third part is about who should hold this capability. We want to make intelligence accessible to individuals and small teams, putting power in people's hands rather than in corporate entities. One of our core beliefs is that intelligence should be accessible to everyone, not just to corporations.

That belief shapes our practice. We run our own work as a one-person company operated with AI agents, so our tools are tested at the scale of one person. We teach people to build their own AI second brain. And we publish guides for small businesses on deciding what to automate and how.

flowchart TD
  V["Vision: AI that thinks alongside people"] --> I["Interface: voice in, concepts out"]
  V --> A["Architecture: structured context, levels, limits respected"]
  V --> P["Access: individuals and small teams"]
  I --> W["The lab's working system and projects"]
  A --> W
  P --> W

Further horizons

Beyond the near work, our plans describe a longer path. Everything in this section is direction, not a built system.

From agents to a cognitive operating layer. The plan runs from single voice agents, to agents that combine voice with memory and reasoning, to VOX OS, our design for an operating layer over agent work, in which an organisation's work is one dependency graph that people and agents share. We call the idea behind this "the corporation as software": an organisation's roles, information flows and decisions specified precisely enough that agents can carry some of them. Our plans describe a continuous model of the organisation's state replacing much of the coordination a management hierarchy does by hand, an idea that the payments company Block set out in its essay "From Hierarchy to Intelligence", and skills that agents can select and combine for themselves. The agentic organisation sets out the design.

Agents with lasting identity. Our plans name coherent artificial personalities: agents that keep relationship context and learn a person's preferences over time, closer to a companion system than a chatbot.

Ethical design built in. Our plans include an ethical framework for agents and safety-first agent design, with an implementation of Isaac Asimov's laws of robotics as the aspiration. Asimov's laws are a literary device, and we treat them as a starting point for thinking about agent governance, not as a specification. The working form of the aspiration is the governance we already practise: written rules about what an agent may decide, checks that stop it when it goes further, and a record that the check fired. Our plans also name transparent AI and human-centred design as aims for every deployment.

Beyond software. We see three further horizons: integration with robotics, research into AI alignment, and guiding the development of AI in directions that improve society. None has a project behind it yet. Our working assumption is that our governance rules, because they are written in terms of authority, evidence and bounded action rather than any one tool, could one day apply to a machine that moves as well as to one that writes files. That is an aspiration, not a capability we have.

The long-term aim

Our long-term aim is to make the world more intelligent through accessible AI, in four areas: better and more accessible education, amplified personal productivity, reducing the damaging effects of technology on society, and ethical AI deployment that benefits humanity. The mission and values page explains the values behind that list, and what we believe about AI sets out the positions it rests on.