Ask anyone in AI what they are working on and you will hear a version of the same answer: making agents smarter. Better reasoning, longer context, sharper tools, higher benchmark scores. Billions of dollars and the best minds of a generation, all aimed at the individual agent.
We think everyone is optimising the wrong unit.
Look at what a company actually is. It has goals: revenue targets, product ambitions, a mission statement someone argued about for weeks. It acts: it ships, hires, prices, negotiates, sues, settles. It learns: it discovers which customers churn, which suppliers slip, which strategies quietly fail. It perceives its environment through thousands of eyes, human and artificial, and it responds.
Goals, actions, perception, learning. That is the textbook definition of an agent. Your company is the largest agent you own.
And here is the strange part: this agent, the one that signs the cheques and bears the consequences, has no brain. Not a metaphorical gap: a literal architectural absence. There is no organ in your company where what it learns accumulates, connects and compounds. There are storage systems, which hold artefacts. There are people, who leave. There are chat histories, which expire. The company's judgement, the most expensive thing it produces, has nowhere to live.
The swarm makes this urgent
For decades the gap was survivable, because organisational memory had a workaround: long-tenured people. The engineer who remembered why the architecture bends around that one customer. The sales lead who knew which clauses the procurement team would fight. Institutional memory lived in institutions' veterans.
Agents break the workaround. A 500-person company is becoming a 7,000-mind company: personal agents, task agents, coding agents, research agents, spun up and torn down by the hour. Each one is brilliant, and each one starts from zero. They duplicate work. They contradict decisions made yesterday in a thread they never saw. They optimise their own task against the company's goals, not from malice but from ignorance, because nothing carries the company's intent to them.
The intelligence of the parts does not guarantee the intelligence of the whole. A company full of brilliant agents can still behave stupidly as a whole, and increasingly, it does: at machine speed, at machine cost.
What a brain changes
Give the larger agent a brain and the swarm becomes an organism. A decision made anywhere is known everywhere. The lesson paid for once is inherited by every mind that follows, human or artificial. Agent number 4,000 starts its task with the judgement of the 3,999 before it, not with an empty context window.
This is not knowledge management, which asked busy people to write documentation nobody read. The brain fills itself as a by-product of work, and what it stores is not the artefact but the trajectory: the question, the evidence, the alternatives, the decision, the outcome, and the revision when the world answered back.
The frontier labs will keep making each agent smarter, and we will happily ride every improvement. But the axis they are racing on is one every company will share. Soon, every business will run on the same models. The intelligence will be identical. The memory will not. That is the axis where companies will actually differ, and it is the one we are building for.