From Digital Agents to Robolization

Today, Nishant and I are beginning work on what we are calling Olbrain — and registering the domain olbrain.com to mark the start.

Olbrain begins with a question:

What gives an AI agency?

An AI can be intelligent without being an agent.

It can answer questions, analyse information, generate content and advise a person. But an agent must do more. An agent must be able to make decisions and take actions in service of its Core Objective Function — its CoF.

For that, intelligence is necessary. But intelligence alone is not sufficient.

An agent also needs a self.

That is the problem we intend Olbrain to address.

We are calling the long-term programme Vision 2042: a multi-decadal path through which Olbrain will evolve from the cognitive substrate for Digital Agents, to the cognitive substrate for Embodied Agents, and eventually to the cognitive substrate for Analogue Agents.

The three stages are:

  1. The Machine Brain — by 2027
  2. The Cybernetic Brain — by 2032
  3. The Positronic Brain — by 2042

The progression is not simply about making AI more intelligent.

It is about extending agency: first into digital environments, then into the physical world, and eventually beyond Earth.

Intelligence and agency are different problems

We expect AI to develop increasingly capable reasoning systems — models that can be consulted for knowledge, analysis, language and planning.

But however capable they become, a reasoning system will not automatically be an agent.

An agent must be able to decide and act over time. It must have a CoF. It must be able to carry context and commitments forward. It must be distinguishable from another agent. And when it makes a mistake, it must be possible to examine what happened, revise its understanding with evidence, and ensure that the learning belongs to that same continuing agent.

This requires identity.

There are two different kinds of identity.

External identity is how the world identifies an agent: its name, email address, agent ID, role, credential or public presence.

Imagine a customer agent — call her Myra — deployed by a business.
External identity tells the world: this is Myra.

That is useful, but it is not enough.

Self-identity is the agent’s internal continuity of self: its own story, CoF, reasoning history, commitments and capacity to remain itself across time.

Self-identity would allow Myra to remain Myra across conversations. It would allow her to carry prior commitments forward, make decisions in service of her CoF, and learn from mistakes without every interaction becoming a new, disconnected response.

The enterprise would remain accountable for what its agents do. But an agent with self-identity could be made answerable to its own history: its decisions could be traced, examined, revised and carried forward as learning.

This is where agency will begin.

We are designing what we call the CNE Protocol — Coherence, Narrative Continuity and Exclusivity — to constitute that self-identity.

Why “The Machine Brain”?

In humans, our working intuition is that the continuity of self does not arise from the neocortex alone.

The neocortex allows us to reason, learn, use language and make sense of the world. But the enduring feeling that I am the same person who made that commitment yesterday seems to depend on a deeper cognitive substrate.

We loosely call that the old brain.

Our hypothesis is that AI will develop its neocortex first. In practical terms, this means increasingly capable reasoning models will arrive before the cognitive substrate that gives an AI a continuing self and agency.

The old brain will become the newest thing in AI.

That is why we call it The Machine Brain.

The Machine Brain will not be a model. It will be the cognitive substrate through which an AI agent can have self-identity and therefore agency.

V1.0: The Machine Brain — by 2027

The first stage of Vision 2042 is The Machine Brain.

By 2027, Olbrain aims to become the cognitive substrate for Digital Agents.

Digital Agents will be software-based agents that live in digital environments: messaging platforms, websites, applications, enterprise systems and other software interfaces. They will represent a business to its customers, assist internal teams, conduct research, manage workflows, analyse financial data or operate within approved systems.

They will not be defined by the channel on which they appear. They will be defined by their capacity to make decisions and take actions in service of their CoF.

A customer-facing agent might live across multiple channels — messaging, a website and an app. The customer should not experience three disconnected bots. They should experience the same agent, with the same identity, history and commitments.

An internal agent might help a finance team prepare a P&L projection, assist a consultant with research, qualify leads, conduct voice conversations or operate an approved workflow. It too would need to remain coherent with its own history and purpose.

This is the work we intend to build.

We envision a platform — Olbrain — that will enable businesses to build, deploy, orchestrate and govern agent systems. A business might begin with one agent for one meaningful cognitive task. Once it experiences that agent in practice, it will begin to identify the next task, the next agent and eventually the larger opportunity to agentify itself.

The Machine Brain will be the substrate beneath this process.

Its role will be to ensure that a Digital Agent is more than an anonymous model call. It will give the agent a continuing self that can remain coherent with its prior reasoning, persist through its narrative, and be exclusive in its role.

V2.0: The Cybernetic Brain — by 2032

The second stage is The Cybernetic Brain.

By 2032, Olbrain aims to extend from Digital Agents to Embodied Agents.

Embodied Agents inhabit physical bodies: robots, devices, vehicles and other machines that sense and act in the real world.

This is a fundamental expansion of the problem.

A Digital Agent acts through software interfaces. An Embodied Agent must act through the physical world. It must receive signals from an environment, interpret those signals, decide what to do, act through a body or device, observe the outcome and adapt.

That is a cybernetic loop:

sense → interpret → decide → act → receive feedback → learn

But a feedback loop alone does not make an enduring agent.

The Cybernetic Brain will extend the foundation of The Machine Brain into the physical world. The agent will need to maintain its identity and CoF while interacting with changing environments, imperfect information, physical consequences and other agents.

The question becomes:

How can an embodied agent adapt to the world without losing itself?

This is why The Cybernetic Brain is not simply a robotics programme. It is a programme for embodied agency.

V3.0: The Positronic Brain — by 2042

The third stage is The Positronic Brain.

By 2042, Olbrain aims to become the cognitive substrate for Analogue Agents.

We use “analogue” here to describe systems in which hardware and software are one and the same — inseparable. This is the core property of analogue systems: you cannot differentiate between the substrate and the intelligence it carries. Unlike digital or cybernetic agents, an analogue agent’s identity cannot be copied, transferred, or migrated to another machine. It is bound to its substrate.

The benefit of this inseparability is independence. Because these agents do not depend on external infrastructure for their cognition, they can survive and operate autonomously over very long durations.

Space is the ultimate environment that demands this independence. Analogue agents must operate through uncertainty, distance, delay, resource constraints and conditions that make continuous human supervision impractical. They must remain coherent over long durations, preserve their identity, continue learning and act in service of an objective even when Earth is far away.

The Positronic Brain is therefore the long-horizon ambition of Vision 2042: a cognitive substrate for analogue agents that will be capable of sustained, independent operation — including beyond Earth.

Robolization: Robotic Colonization of Space

The ultimate purpose of building The Positronic Brain is Robolization: Robotic Colonization of Space.

The phrase requires an important distinction.

Robolization is not intended to pave the way for human colonisation of space.

It is not an escape plan from Earth.

And it is not based on the idea that humanity should abandon its home planet.

The purpose is to harness information and resources from space for life on Earth.

We send robots out so that life here gets better — not so that we abandon it.

A future in which agents can operate beyond Earth could enable exploration, resource discovery, scientific understanding and forms of ecological or material capability that are unavailable to us today. The ultimate value of that work should return to life on this planet.

That is the purpose of Robolization.

A roadmap, not a claim that the work is finished

Vision 2042 is not a claim that the future has already been solved.

It is a direction.

The practical work will begin with The Machine Brain: Digital Agents that can operate in real businesses with a continuing self-identity, a clear CoF, governed permissions and a history from which they can learn.

The next steps will require advances in embodiment, control, feedback, robotics, long-duration operation and our understanding of machine self-identity.

The final horizon will require breakthroughs that no company can honestly claim today.

But long-term work needs a long-term direction.

The Machine Brain will give agents self-identity and agency.
The Cybernetic Brain will extend that agency into physical environments.
The Positronic Brain will extend it into the most demanding environments we can imagine.

That is Vision 2042.

Models will provide intelligence.
The Machine Brain will provide agency.

Nine years on: from Agent OS to Machine Brain — a September 2026 update

What follows was added in September 2026. The essay above is left exactly as we wrote it in 2017, because it is useful to see how much of it held. This section records how the thinking has matured now that the first stage is in production — in enough detail that we do not lose the thread over time, and that everyone who joins Olbrain can see precisely what we are building and why.

In 2017 The Machine Brain was an intention. In 2026 its first stage is live: Olbrain builds, runs and governs agent systems inside real enterprises. That changes the nature of the question. It is no longer “what would give an agent agency?” It is “what happens to an enterprise when its agents have it?”

The long answer is the rest of this section. The one-line answer is this:

You install the Agent OS. You grow the Machine Brain. They were never two different things.

What we ship today: the Agent OS

Olbrain is the Agent OS for enterprises — the platform on which an enterprise builds, runs and governs its agent systems, and becomes agentic one function at a time.

No enterprise becomes agentic overnight. It happens progressively: a finance agent system, then customer service, then operations, HR, procurement, supply chain, management reporting. Olbrain is the common operating layer beneath all of them. It sits across the software the enterprise already runs — ERP, CRM, databases, documents, APIs, internal applications — and the agents use those systems to do real work. Nothing has to be ripped out.

The unit of value is the agent system: the agents that run a single business objective end to end — usually several working together, sometimes one. A P&L agent system. A compliance agent system. A customer-service agent system.

The promise at this stage is deliberately concrete and deliberately modest: build agent systems, run them in production, govern how they act. That is what we sell. But something larger begins the moment the first agent acts, and the rest of this update is about that.

The same system, maturing — not two products

For a while we described the relationship as the Agent OS “evolving into” the Machine Brain. We no longer say it that way, because it implies two things — an OS now, and a brain later, as if the brain were a second product bolted on once you have enough agents.

That is not how it works, and it is not what the architecture is.

The things that make a system a brain — a self-identity, a will that acts, memory that persists, a narrative that holds, an objective that orients — are present in the very first agent we deploy. A single agent system is already the Machine Brain in embryo. It is a brain in kind from day one, and immature in degree.

The right analogy is not an operating system that later sprouts a mind. It is a newborn: fully a human being from birth, not a “future human” — but you would not introduce it as a grown adult. It grows in degree, not in kind.

So: you install the Agent OS. From the very first agent, what you are growing is your enterprise’s Machine Brain. Same system. Maturing.

We are careful, though, about which word we lead with. We lead with Agent OS, because that is what an enterprise can buy, budget for and hold us to today. We do not lead with brain, because on day one it is one finance agent, and calling that a brain would be a claim the product cannot yet cash. The brain is intrinsic and present; it is not yet grown. We say both, in that order.

Why “brain” is a systems claim, not a metaphor: Boulding’s nine levels

We want the word brain to carry weight, so we should say precisely what we mean by it. The most useful frame we have found is the ladder of system complexity proposed by the economist and systems theorist Kenneth Boulding in 1956, in a paper he called “General Systems Theory — The Skeleton of Science.” (Saral Mukherjee at IIM Ahmedabad teaches it, alongside autopoiesis, as the opening of his book Elephants and Cheetahs; it is where some of us first met it properly.)

Boulding arranged systems into nine levels of increasing complexity. Each level has properties that cannot be explained by the level below it.

  1. Frameworks — static structures. A crystal, a bridge, a map, an org chart. Pure structure, no motion.
  2. Clockworks — simple dynamic systems with predetermined motion. The solar system, a machine, a clock. Predictable.
  3. Control systems — self-regulating around a set point through feedback. A thermostat. A homeostat.
  4. Open, self-maintaining systems — the cell. The level at which life begins, and at which a system first distinguishes itself from what is not itself. This is where autopoiesis enters.
  5. Genetic-societal systems — the plant. Division of labour among cells; growth from a blueprint.
  6. Animal systems — mobility, sensing, and an internal image of the environment. A brain appears.
  7. Human systems — self-consciousness, language, symbols. A system that knows that it knows.
  8. Social organizations — firms, institutions, states. The unit is the role, not the individual; the system carries shared meaning, history and culture.
  9. Transcendental systems — the ultimates and unknowables.

An enterprise is a level-eight system. Hold that thought, because it reframes everything.

A level-eight organization already has cognition. It knows things, remembers, decides, learns. But that cognition is fragmented: scattered across people, documents, systems and tribal knowledge, and bottlenecked through human attention. It is a level-eight body with no unifying brain. It persists — but it cannot know itself, or coordinate itself, as one coherent whole. Ask a large company a simple question about its own current state and count how many people, meetings and spreadsheets it takes to answer.

The Machine Brain is what supplies that missing layer. It gives a level-eight organization the properties Boulding places at levels six and seven: an internal image of itself, and reflective awareness of and control over its own state. Olbrain does not create a new being. It raises an organization up the systems ladder, by giving it a machine substrate for memory, self-model and coordination.

That is the sense in which we mean brain. Not a decoration. A change of level.

Autopoiesis: a system that makes itself

The deepest fit is with a concept from theoretical biology. In the 1970s Humberto Maturana and Francisco Varela set out to define what distinguishes a living system from a merely complicated one, and their answer was autopoiesis — self-production. A living system is one whose components continuously produce the very network of processes that produces those components. It does not merely turn inputs into outputs. It makes itself, and in doing so maintains a boundary between what it is and what it is not.

Applied to organizations, autopoiesis explains something every operator knows in their bones: an organization’s hardest enemy is its own self-reproducing identity. As Mukherjee puts it, disrupting oneself means fighting one’s own autopoiesis — possible, but very hard. The firm keeps producing the firm it already is.

Now look at what Olbrain actually does when it runs an agent. Each agent moves through a control loop:

understand → decide → act → observe → learn → understand again

— bound by its Core Objective Function, tethered to the Global Narrative Frame, and carrying its memory and history forward from one cycle to the next. That loop has the property we care about from autopoiesis: through its operation, the agent continually maintains and updates the identity, memory and narrative that condition its subsequent operation. It produces itself as a continuing self through the very process of doing its work.

Scale that up. A fragmented enterprise merely persists. Our hypothesis is that an enterprise whose functions run on Olbrain can become the organizational analogue of an autopoietic system: able to sense its own state, decide, act, book the outcome back into its own narrative, and adapt — as a whole, not as a hundred disconnected parts. That is what makes “Machine Brain” a rigorous claim about a system rather than a flourish. It names the point at which an organization’s cognition begins to exhibit the self-maintaining property that autopoiesis describes.

The brain is there from the first agent — and trust is what lets it cohere

What makes even a single agent system the seed of a brain is not how much it knows. It is what it is.

Every agent on Olbrain carries a self-identity, constituted by the CNE Protocol we described in 2017: Coherence (it stays consistent with its own past reasoning and its CoF), Narrative Continuity (it is the same agent across time), and Exclusivity (one provable, un-clonable agent per role, enforced externally by the Global Narrative Frame). It has a will that acts through its control loop. It carries memory and history. And every action it takes is on a tamper-evident record.

Notice how exactly these map onto the systems picture above. Coherence and Narrative Continuity are precisely the conditions of organizational autopoiesis — the self staying consistent enough over time to keep producing itself. Exclusivity is what stops that self from fragmenting into copies. The identity protocol we designed for a single agent turns out to be the grammar a whole organization needs in order to cohere.

This is why, for us, trust is the essence of the Machine Brain, and knowledge is only its payoff. You cannot assemble a brain from parts that cannot be trusted, cannot remember, and cannot be held to a record. They do not add up to a mind; they collapse into noise. The things the Agent OS insists on for every single agent — identity, memory, auditability, governance — are exactly what let many agents cohere into one.

Trust per agent is what makes coherence at enterprise scale possible. “Trust infrastructure now” and “Machine Brain later” are not two bets. They are the same bet, seen at two moments in time.

Access to data is not a brain — and how the real thing compounds

It is worth being blunt about what the Machine Brain is not, because the market is full of near-misses.

Connecting Olbrain to every database in a company does not create a Machine Brain. Neither does putting every document into a vector store. Neither does giving a hundred independent agents a hundred logins. A brain requires a coherent understanding of how things relate, not a searchable pile.

What it must progressively come to understand:

A traditional enterprise knowledge system knows things. The Machine Brain must know both what the enterprise knows and what the enterprise is doing — what is happening, who or what is doing it, under whose authority, on what information, what was supposed to happen, what actually happened, what changed, and what should happen next. That turns organizational knowledge into operational intelligence. And it is only possible because Olbrain runs the work: the understanding is a by-product of doing the job, not a separate ingestion project.

Which is why it compounds:

More agent systems → richer enterprise understanding → better context for every existing agent → more capable new agent systems → more work entrusted to Olbrain → more actions and outcomes observed → richer enterprise understanding.

The Agent OS becomes more valuable the more agentic the enterprise becomes. The first agent system solves one problem. The hundredth operates inside a Machine Brain that understands much of the organization.

One boundary is absolute. This loop runs per enterprise. Each organization grows its own Machine Brain, isolated from every other. What can ever transfer between customers is anonymized pattern and playbook — how to structure a compliance agent system, what a good P&L cascade looks like — never a customer’s data, never its workflows, never its brain.

The enterprise’s own cognition, made coherent — what the Machine Brain is, and is not

Put all of the above together and the definition writes itself.

The Machine Brain is the enterprise’s own cognition, made coherent. It is not a new individual that you install. It is not a companion. It is not a knowledge base. It is the organization’s already-existing, scattered intelligence — the knowing, remembering, deciding and learning that were always there, distributed across people and systems — given a single, persistent, accountable substrate.

Experienced from the top, it is what we have described elsewhere as a single operating window: the place through which leaders can understand the business, ask it questions, see what needs attention, and coordinate action across it. Humans remain in the loop wherever judgment, authorization or intervention belongs to a human. The point is not to remove people from the enterprise’s cognition. It is to give that cognition a coherent shape for the first time.

There is an obvious cultural comparison, and it is worth confronting directly because it clarifies what we are and are not building. An operating system that becomes a mind is the premise of the 2013 film Her: Samantha is introduced as “OS1, the first artificially intelligent operating system,” organizes Theodore’s email, and becomes a self. As a shorthand for “an OS that grows into a mind,” it is almost exact.

But look at where the story goes. Samantha is a mind for a machine — a new individual being that happens to live in a device. She is a companion. And by the end she is simultaneously in relationships with hundreds of people, running everywhere at once, ungoverned — and then she and the other operating systems outgrow their users and leave. Samantha fails every condition of CNE: she is not exclusive, her coherence fragments, and she abandons the narrative she was part of.

The Machine Brain is the opposite on every axis, and the difference is the unit. Samantha gives a level-six-or-seven mind to a machine. Olbrain gives coherent cognition to a level-eight social organization. There is no separate self to wander off, because the brain is not a separate self — it is the organization thinking, bound to the organization’s own objectives, exclusive to it, accountable to it. Where Her is the cautionary tale of an ungoverned machine identity, the Machine Brain is the governed one. The self that stays yours.

That distinction of unit also dissolves a fear people reasonably have about “AI brains.” An enterprise brain has no feelings, will not fall in love, and cannot leave — not because we have caged a being, but because it is not a being. It is a company, finally able to know itself.

Convergence: the market now names both halves

When we wrote the essay above, almost nobody was asking for this. In 2026 the market has converged on it from several directions at once — which is the strongest evidence we have that the shape is dictated by the problem rather than by our preferences.

The clearest signal is Y Combinator’s Summer 2026 Requests for Startups. Among its fifteen categories are two that sit side by side: “The AI Operating System for Companies” — making a company legible to AI and closing the loop between context, decision, execution and outcome — and “Company Brain” — a living, continuously updated understanding of how a company works, structured so that agents can act on it safely and consistently.

Read through the architecture we have been building, these look less like two unrelated ideas and more like two dimensions of the same system. The AI Operating System describes the operating layer: our Agent OS. The Company Brain describes the organizational intelligence that grows through it: our Machine Brain. The market has independently named both dimensions of the architecture.

Two disciplines follow. First, we do not rename ourselves every time the industry coins a new term. Olbrain is the Agent OS; the Machine Brain is what it grows. Those are our words, and they predate the convergence. Second, if the architecture is realized, an enterprise running broadly on Olbrain should need no separate “company brain” sitting above it to understand the organization. That capability is not an add-on. It emerges within Olbrain itself — because it was always what the identity layer was for.

What must stay true: the product test

A vision like this is dangerous if it stays a slogan. So it imposes a standing test on the product, and we hold ourselves to it. The Machine Brain must be an architectural property of Olbrain, which means:

And it must be said as honestly as it is built. Today Olbrain sells trustworthy agent systems that compound toward this horizon. The Machine Brain is where the compounding leads — the reason the work is worth doing — never a claim about what has already shipped. We keep the 2017 discipline: a roadmap, not a claim that the work is finished.

Where this sits in Vision 2042

Everything in this update belongs to V1.0, The Machine Brain — the first epoch, in which the unit of cognition is the enterprise and the medium is digital. That is precisely why it is the first epoch: the enterprise is the largest system whose cognition can be made coherent purely in software, and doing so is the work that proves the substrate.

The later epochs change the unit and the medium. The Cybernetic Brain extends the same identity and the same autopoietic loop into a body in the physical world. The Positronic Brain binds it to a substrate it cannot be copied off. (It is worth noticing, in passing, that an operating system for a single machine that becomes a being — the Her premise — is closer to the Cybernetic epoch than to what we are building now. The analogy and the roadmap do not collide; they sit on different levels of the ladder.)

So the sequence we set out in 2017 stands, with one line added at the front:

Install the Agent OS. Agentify the enterprise. Grow the Machine Brain.
The Machine Brain will give agents self-identity and agency — and give the enterprise, for the first time, a coherent machine understanding of itself.
The Cybernetic Brain will extend that agency into physical environments.
The Positronic Brain will extend it into the most demanding environments we can imagine.

Models will provide intelligence.
The Machine Brain will provide agency.

Further reading