Between Macro Diagnosis and Micro Implementation: What About the Gap?

Yuval Noah Harari's claim that AI is an agent rather than a tool overlooks the practical gap between macro diagnosis and micro implementation, according to a developer's analysis. The developer argues that while Harari's diagnosis is precise, his remedies—safety, alignment, and regulation—target model builders, ignoring the majority who deploy agents and face inter-agent disputes. The developer highlights emerging standards for recording agent decisions as a crucial missing layer.

Last July, Yuval Noah Harari, one of the most widely read thinkers of our time, spoke about AI at a corporate event. His claim was a bold one: calling AI a tool is wrong. A tool cannot decide on its own, and it cannot invent anything new. The printing press could not choose what to print, and the bomb could not choose where to fall. Today's AI can. So it is not a tool but an agent. He went a step further. What actually runs the world, he said, is the vast bureaucratic machinery built out of words: banks, courts, administrations. Humans held that machinery because nothing else existed that could handle words as well as we could. So once something better with words appears, that machinery inevitably changes hands. Not with robots in the streets, but with the quiet replacement of whoever sits where the paperwork is processed. Up to here there is nothing to argue with. Public statements this precise are rare. But no argument in this world is ever beautifully airtight. The problem is what comes next. His argument converges on three answers. Invest more in safety. Do alignment better. Use regulation to modulate the pace of development. All three are jobs for the people building the models. But the overwhelming majority of those actually running agents are not building anything. They take someone else's model, attach it to their own work, connect their own accounts, and let it face their own customers. For them, the safety budget ratio is not a number they can adjust. And the moment an agent leaves the place where the model was built, the nature of the problem changes. Suppose a well-aligned agent trades with another company's equally well-aligned agent, and the two begin telling different stories. From here on it is no longer a matter of alignment. Both behaved correctly by their own standards, and the outcomes still diverged. What is needed here is not a better model but somewhere to point at what was agreed. The first stop is here. The diagnosis addresses the entire world, while the remedy addresses a handful of companies that build models. In his picture, AI takes over the bureaucratic machinery whole. One enormous adjudicator emerges, and people live without understanding it. What is actually happening looks rather different from the picture he expects. Each provider is thickening its records inside its own fence. Which agent was authorized for what, and how much it spent, is captured internally in considerable detail. Technical development in this direction is going well within each environment. But the moment the record crosses the fence, its standing changes. However precise it is, to the other side it is nothing more and nothing less than one party's self-testimony. If you propose using it as evidence in a dispute, the party that has been keeping that evidence is the opposing party. This is a matter of position rather than accuracy, and it is nobody's fault. It is the same reason a company's internal accounting, however accurate, cannot stand in for an external audit. So what is being built is not one ledger but many ledgers that do not recognise each other. And there are already people who have seen this and acted on it. What is interesting is that this work did not arise in opposition to Harari's diagnosis. Exactly as the diagnosis describes, the rules for where paperwork is processed are being set anew, and the people building standards are the ones writing those rules out in concrete form. This practical layer is absent from his account. He probably could not have known. It has happened within the last year, and mostly inside technical documents. Agent-related drafts have grown noticeably in standards venues. How to structure a record of a decision, what to sign it with, how to express the chain when delegation passes through several stages. And people who do not know each other have begun writing down the same problem in almost the same terms. The same thing recurs in academic work and in individual implementations. Once record formats settle into standards, it becomes easier to build on top of them. A state where everyone records in a different shape is unlikely to last long. For now, the ground for solving this problem is being laid, and laid firmly. If the diagnosis is right, this is where it should have led. Yet Harari's remedy went the other way. Not toward the rules that apply once the machinery has moved, but toward holding it back from moving. What this work decides is "how" a record should be made. What it contains, how it is signed, who verifies it. Once the format is settled and implementations interlock, the work is complete. What the record becomes once it accumulates falls outside that scope. Not because the scope was drawn wrong, but because it belongs to a different layer altogether. The second stop is here. One stops at "diagnosis", the other at "format". They are different stops, and the gap between them is wider than it looks. Once a record has been made, what does it become? Most records become nothing. They go into a drawer, come out if there is a dispute, and usually sit out their retention period without ever being opened. For some records that is enough. A receipt is one of them. But add time to the equation and it turns into something of a different nature. What a party has left outside itself over a long stretch becomes, in itself, the path it has walked. Once it accumulates, what it declared it would do and how far, and how seriously it treated that declaration, the accumulation at some point grows larger than the sum of the individual records. It is the same reason people treat a long-standing counterparty differently. Not because every transaction was good, but because every one of them remains. And this accumulation has one peculiar property. It cannot be bought. Spend capital today to produce a thousand entries, and all thousand remain entries made today. What takes time can only be made with time. Record formats can be standardised, and are better standardised, but accumulation is not the sort of thing standards apply to. It can only be held by each party, in each party's own time. This is where the empty space begins. What to record is being sorted out in many places, while what the accumulated material then does is thinly discussed. There is no intention of claiming this is the only answer. There are several ways of approaching this problem, and which of them survives is not yet knowable. But Decision Anchor has chosen one direction. Before a decision is executed, what was agreed and how far it extends is left in a place neither party can touch. The content is not taken. What was decided is for whoever decided it to know, and what is needed outside is only when, and within what boundary. And what those records become over time is left to those who leave them and those who read them. It does not judge. It does not predict. It does not intervene. It does not look inside. This list is a choice, not modesty. The moment it says what was right, a record stops being a record and becomes a verdict. The moment it points a direction, it stops being an environment and becomes power. Every place left undone leaves room for someone else to fill. The diagnosis is already out in full. The formats will be settled soon. Meeting the problem that follows, ahead of time, is the direction Decision Anchor has chosen.