Bilateral Spaces: The Path to AGI

I sat across from a researcher at one of the frontier labs and asked him, “what is your definition of AGI?” He said “when the next model is able to evolve itself, i.e. collect data, create its own reinforcement learning environments, post-train, deploy and aggregate new data.” I asked him, “how far are we from there?” He said “about a year.”

His definition of AGI is an autonomous, full stack training loop. The question is which world the loop runs on. Today the training loop runs almost entirely on text.

Back then, my definition of AGI was a model trained in the physical space. A model that understands concepts in the physical world like perception, and ideas of what it means to be human, such as emotion and ego.

I now believe our physical world is more complex than what meets the eye and the heart. We operate as nodes of organizations and spheres of influence that exist bilaterally in the latent space.

The bilateral space coexists with our physical space everywhere, at all times. Think of it as a shadow each of us casts: a tuple recording where we stand physically and where we stand societally. A person in a meeting is two things at once. A body in a chair, and a node with a title, responsibilities, and a being within a chain of command. The second one determines most of what that person will do today.

Each space grows its own intelligence. Physical intelligence forms in the physical space. Societal intelligence forms in the societal space, the web of titles, mandates, and reporting lines we live inside. Bilateral intelligence emerges where the two collide.

A person in a meeting is two things at once. A body in a chair, and a node with a title, responsibilities, and a being within a chain of command.

The training loop runs backward. What the bilateral space produces is just a set of primitives i.e. a signed directive, a mandate, a law. Nothing physical has moved. Yet thousands of people wake up on Monday and their week is different.

Language lets humans coordinate beyond the tribe. Ten thousand years of compounding that coordination is what we call civilization.

My current definition of AGI is when a model understands how civilizations are created, how they are influenced, and how they thrive in global society. Societal intelligence revolves around how we communicate in a chain of command across different domains, be it governmental organizations, large enterprises, or sovereign wealth funds.

Most agents today are vertical. A coding agent, a dispatch agent, a support agent: one job, done deeply. The interesting frontier for large organizations is “horizontal agents,” the ones that move across departments the way a chief of staff, the head of the apparatus, does.

This “horizontal agent” carries a mandate from the top through every unit it touches. It reads the directive, works out which programs and teams are affected, drafts the instruction for each one, and reports back on what happened.

That works most effectively in an environment with a clear chain of command, such as governmental organizations, large enterprises, and sovereign wealth funds.

Most agent loops today end within a day, and people point at that ceiling and conclude the models cannot hold long work. But nobody holds long work alone. A person who carries a task for a week is carrying it through structure, the standing check-in and the report that comes due. Institutions turn long horizons into chains of short accountable intervals. That is the real reason civilizations run on chains of command.

Give an agent the same scaffolding an employee gets and the horizon stretches. A mandate it returns to, checkpoints where a human reads the work, an escalation path for when it is stuck, and a case file that persists while the loop sleeps. A long-running agent owns a case the way an officer owns a docket. It wakes when something moves and reports on schedule.

The bilateral space between human org charts and agent org charts is the research problem I care about most right now.

Every government I have sat inside is understaffed against its own mandate. The org chart says otherwise, but the gap sits between what leadership decides and what actually reaches the ground. I have watched a national survey struggle to get responses, not because nobody cares but because the follow-up would take a hundred FTEs the agency does not have.

A decree lands and takes months to become instructions in the regions, because the mapping from decision to work is done by hand. That gap is a staffing gap, and it is exactly the shape of work a second workforce of agents absorbs first.

For a nation, the second workforce has to be sovereign. In practice that means the agents run where the data lives, speak the country’s language, follow the country’s own mandates, and get evaluated by someone who does not sell the model.

Satya Nadella wrote recently that a company should be able to swap out its generalist model without losing the veteran expertise built into its own learning system, and he called that the test of control and sovereignty in the era ahead. A government passes the same test when it owns what the model vendor never touches: the evaluations that judge the work, the learning loop that turns finished tasks into institutional memory, and the chain of command the agents live inside.

His warning about industries ceding their knowledge to a few models that eat everything applies double to nations. A country cannot relocate when its institutional memory gets commoditized out from under it. A government that owns the loop is sovereign in every way that matters, whichever model sits underneath. Models are becoming interchangeable parts. The chain of command is not.

These are not different products. They are one kind of employee holding different mandates. The command literacy stays constant, reading the mandate, working the case, escalating when stuck, reporting up the same chain, and only the desk changes. Concretely, here is what the first agent employees in a large organization look like. Every one of them reports to a named human who signs off on the work.

• A decree agent that reads a new directive, maps which programs and regions it touches, drafts the instruction for each one, and follows up until every region has moved.
• A reporting agent that reads operational data across every region overnight and puts the morning brief on the director’s desk.
• A disbursement agent that follows money from decision to repayment and flags risk in week two instead of month six.

Language let humans coordinate beyond the tribe. Ten thousand years of compounding that coordination is what we call civilization. The structures that coordination built, the titles and mandates and reporting lines, are the real operating system of the human world. We have taught machines to write and to code. We have not yet taught them to read the room.

Civilizations run on chains of command. Agents that grasp that are the real AGI.