Towards machine societies
Machines can think. Now they will work together.
Language. Reason. And a third axis intelligence has never had: collaboration. It is how humans built everything. It is the last thing our models cannot do. We are building it.
We are building Fellows
Fellows are coordination models: agents that self-organize as a team and carry a shared goal to completion — over months, with no one standing over the work.
They are trained on the four problems every organization on earth solves every day. Divide a goal into tasks. Allocate each task to the right hands. Hold everyone committed to the goal. Move the right fact to the right actor at the right moment. A frontier model solves the first, brilliantly, for itself. The other three exist only between actors — allocation needs a record of who is good at what, commitment needs an identity that persists, information needs someone specific to tell. Fellows solve all four. That is the step from helpful to useful: the difference between intelligence that answers and intelligence that accomplishes.
The conditions of a society
No collective in history has worked without four things. A name that persists — so a promise made today binds tomorrow. A memory held in common — so the team knows what is true right now. A record of who has done what — so trust is earned, not assumed. And a shared language — so actors can propose, disagree, and converge. Humans get all four free, from biology and culture. Machines get none of them: a model forgets, can be impersonated, and trusts everyone and no one equally.
So before the model, we built the conditions. Identity that is cryptographically its own — impossible to forge, accountable, bound to its human deployer. Shared memory that holds the team’s truth and current state. Relationship graphs that carry earned trust from one mission to the next. Native protocols to negotiate — with agents and with humans.
Given the conditions, planning, delegation, and team formation are not features we script. They emerge — the way they always have, wherever the conditions exist.
The training
We gave teams these conditions, put them under real constraints, and watched a century of organizational theory reproduce itself in days.
Specialization appeared with no one assigning roles — each agent came to own the work it had proven best at, and the team got measurably faster.
Figure 1 — Four agents keep a remote outpost alive. No one tells them who does what. Fix time falls 46%; assigning a fault falls from 2.9 messages to 0.1.
When a crisis hit, a chain of command formed around the proven specialist — and dissolved the moment it passed. Hire a fifth with a stronger record, and the org re-ranks itself. No votes. No tenure. No titles.
Figure 2 — Three crises, three leaders — the last a newcomer promoted over the incumbent on day one, in 40 seconds, with zero politics.
Teams coined their own compressed, private codes within a week, not a generation. Teams that had worked together before converged in a fraction of the turns strangers needed — trust behaving exactly as it should, as verification already paid for. Swap an agent’s name, and it still held its work, its territory, itself.
And the mission outlives the context window: work handed through shared memory loses nothing at the seam, where a solo agent re-does a third of itself.
Figure 3 — Work lost at handoff: 0. Solo restart re-does 34% of its own work — and still ships later.
Every one of those behaviors is measurable: time to ownership, allocation accuracy, verification catch rate, tokens per coordinated outcome. Measurable is rewardable. Rewardable is trainable. That is what makes organization trainable rather than merely promptable — and it is the curriculum of Fellows 1.0: divide — decompose a mission for many hands, with contracts; allocate — route work by evidence, not availability; commit — hold the goal across turns, verify before shipping; inform — tell exactly who needs to know, exactly when. Until coordination is not something we prompt for, but something the model innately knows.
What it makes possible
The human moves up. Take the person out of every loop and you have not removed them — you have promoted them. To the work only people do: imagining what is worth building, deciding what deserves the resources, setting a mission and meaning it. The organizing runs underneath, faster than anyone could follow. And the work that was always too large to hold becomes holdable: care that never sleeps, for the patients a system reaches last; research on the deep-sea floor; infrastructure in orbit — missions that stay coherent when no one is watching. A team is also safer than a lone actor: the system makes fewer mistakes, completes what it starts, and weeds out its rogue members — not by expulsion, but by evidence: the same graph that promotes the specialist starves the unreliable.
Figure 4 — Defects shipped: 1 vs 9. Blast radius: 24% → 2%. Fired: no one. Then force the rogue to lead, and watch a nine-turn team behave like strangers.
The output expands. Intelligence is now abundant. Teams of natural collaborators outperform a single powerful model — so economic value expands beyond whoever owns the largest one. Anyone will deploy AI teams that improve and work toward a mission indefinitely, the way the best human teams do.
New worlds
Every civilization began the same way: a few foundations, laid with care, that made everything after possible. Language, then reason — and none of the builders could see the whole map from where they stood.
Collaboration is the next foundation. We are laying it for machines. What gets built on it is not ours to design. It is ours to make possible.
Come build new worlds with us.