Fifty doublings

Evolution gave us linear intuition. One day's work, one day's result. AI demands exponential reasoning, because one agent can become a thousand in a minute. The hard part isn't the math. It's unlearning our instincts.

Here is what the curve looks like from inside a company. In 1865 William Jevons noticed that a more efficient steam engine didn't save coal. Britain burned more, because cheap power kept finding new uses. Satya Nadella made the same point about AI: as intelligence gets cheap, we use far more of it, in far more places. Cheaper answers mean more agents doing more work, every year. One good decision can now run a thousand times before lunch.

A company is a set of business loops, and it makes money by turning the same loops faster and better each time. The deal loop: discovery, champion, procurement, close. The renewal loop: onboarding, tickets, escalation, signature. The incident loop: alert, diagnosis, fix, postmortem. None of them runs alone. Go-to-market feeds engineering, engineering feeds product, product feeds every deal and renewal, and the field's lessons come back around to the next plan. The most famous loop in business is the one Jeff Bezos sketched for Amazon: lower prices bring customers, customers bring sellers and selection, scale lowers prices again. It still turns because every turn made the next one easier. Every turn of a loop teaches the company something. Today the lesson lands in one agent's context or one person's notes, and the next turn starts without it.

Here is where the curve goes. Within a few years agents stop assisting with steps and start running loops end to end, with people setting the goal and handling the exceptions. Agents outnumber people, first in a few functions, then across the company. By then it matters less how smart each agent is than whether the thousandth turn of a loop is better than the first. Ten years out, the winning company is a set of loops that improve themselves, running on a shared memory of what it decided, what it tried, what worked and why. That memory is the system of record for how the company thinks, the way the database was the system of record for what it did. Models get swapped like hardware. New agents join by reading it. Unfold any origami model and every crease is still in the paper; with the pattern, the shape can be folded again. The memory is the crease pattern. It's what compounds, and it's the part the company owns.

Clayton Christensen's rule: products stay integrated while they aren't good enough, and modularize once they are. Today what a company's agents learn lives inside each agent, tied to that agent's model and vendor, and the company changes both often. As agents get good, that learning becomes a layer the company owns, readable by every agent and every model it runs. That layer is what Anona Labs builds. We are a company built for agents: we work out what agents need to get better on every run, and we build it.

It is one system with three parts: memory, review, replay. One shared memory that every agent reads and writes, with the evidence attached and the history kept, so a rule that held in March and broke in June shows both dates, and nothing the company learned is lost when a model is swapped or an agent is retired. A review at the end of every run, where the agent that ran it writes down what it did, what it got wrong, and why, and the review goes into memory so the next run starts from it. The agent holds the best raw material for that review; the outcome supplies the verdict. And replay, the part biology found first. It spent a few hundred million years learning that memory isn't finished while you're awake. It's finished afterwards, in sleep. In 1994 Matthew Wilson and Bruce McNaughton recorded the brains of sleeping rats and watched them rerun that afternoon's maze. The replay is how a brain keeps what mattered and drops the rest. Agents don't sleep, so replay runs alongside them, all the time, and it reads across the loops: the incident at 03:00 becomes a line in the renewal brief in June, the objection that killed three deals becomes a requirement engineering ships. Each loop learns from its own turns. Replay is where the loops learn from each other, and the memory gets sharper, not just bigger.

Run, review, replay, run again, better than last time. That is the loop under all the other loops.

What people do in that company is what we care about most. Machines take the repetitive shift. People set the goals and decide what is worth doing. Most of us will get to be beginners again, and that's the good part.

An origami crane starts as one flat square. No cuts, no glue. Every fold is made on the fold before it, and the paper keeps every crease. A company whose agents run its loops becomes the same thing: thousands of hands folding one sheet, and the sheet remembering. Fifty folds take you to the sun. Not one of them works without the one before it. That is what memory is for.