Unison Without Agreement

AI agents will not need to conspire to move together. They only need to be alike.

The average American checking account pays about 0.1 per cent interest. A few taps away, accounts at online banks and fintech firms pay between 3.3 and 5 per cent. That gap has survived for years because most people do not move their money. They are busy, or loyal out of habit, or they cannot face the forms. Banks have quietly built their business on this. Cheap deposits that stay put fund the loans.

Torsten Slok, the chief economist at Apollo, has asked what happens when every household has an AI agent whose job is to notice that gap. Assistants like Meta’s new Muse, he suggests, could soon sweep household cash automatically into the better-paying accounts. The agent is never busy, has no habits, and does not mind forms. It moves the cash. Repeat that across a hundred million households and the banks lose the cheap funding their lending depends on.

Lazy money was a safety feature. Nobody wrote it down, because nobody had to.

A run at machine speed

Slok is describing a slow bleed. A slow bleed and a stampede are closer than they look. When Wachovia failed in 2008, about $10 billion left it over eight days. On 9 March 2023, depositors pulled more than $40 billion out of Silicon Valley Bank in a single day, and the bank expected to lose more than $100 billion the next. Regulators closed it the following day. That run moved at the speed of phones and group chats.

The friction left in it was human. Somebody had to hear the rumour, believe it, and open the app.

An agent removes that last friction too. The same agent that moves your savings for three to five extra points of interest can move them within seconds when a credible rumour appears, and millions of its siblings may be reading the same rumour at the same moment. Deposits that no longer stick make a bank less profitable and easier to run, both at once. The slow bleed and the fast run are one problem seen at two speeds.

Unison without agreement

Gary Gensler described the mechanism underneath in 2020, while he was still at MIT and before he chaired the SEC. In a working paper with Lily Bailey, he argued that deep learning in finance would tend towards uniformity, because a handful of data aggregators and AI providers would supply much of the data and many of the models. “Models built on the same datasets,” they wrote, “are likely to generate highly correlated predictions that proceed in lockstep, causing crowding and herding.” The paper named a quieter risk as well, one its authors judged less evident than the others for now: that “advice provided by each virtual assistant becomes standardized and commoditized, causing herding of client decision making.”

As SEC chair, Gensler repeated the warning in 2023: AI “could promote herding with individual actors making similar decisions because they are getting the same signal from a base model or data aggregator.” A base model is the large general-purpose system, trained at great expense by a few companies, on which thousands of specific products are built.

This is sometimes described as agents coordinating “acausally”, reasoning about one another and moving together without ever exchanging a message. That can happen, but most of what is coming is simpler. It is correlation from a common cause. When the station clock reaches five past eight, a thousand commuters who each decided alone to catch the 8:10 all start for the same platform. None of them arranged to meet, and the clock did not tell them to. It only gave all of them the same moment. Agents built on the same few base models, fed the same data and told to maximise the same yield will reach the same answer at the same time. No message passes between them, and none needs to.

The distinction matters because it changes the remedy. A conspiracy is broken up by cutting its channels. A common cause is broken up by diversity, which means different models, different data and different objectives, and by keeping any one provider from becoming the clock that everyone reads.

When smarter makes it worse

It is tempting to assume that more capable agents will be wiser ones, and that capability will cure the herding it creates. Whether it does depends on the kind of game being played.

In some games, what others do pushes you the same way. A bank run is the classic case. Douglas Diamond and Philip Dybvig showed in 1983 that a perfectly sound bank can fail if enough depositors fear that others will withdraw, because each withdrawal makes the next one more sensible. In a game like that, an agent that predicts other agents better reaches the run sooner. Capability accelerates the crowd.

In other games, what others do pushes you the opposite way. A crowded trade is the mirror image. The more agents buy the same underpriced asset, the less profit remains for each. A capable agent that sees the crowd forming goes elsewhere. Here capability disperses the crowd. Economists call the first kind of game one of strategic complements, because moves reinforce each other, and the second one of strategic substitutes, because moves stand in for each other.

Personal finance is also more personal than the word “optimal” suggests. Tax positions, debts and time horizons differ, so the best portfolio for one household is not the best for the next. The games to watch are those in which following the crowd pays, played by agents that all learned to think in the same place.

The same pattern beyond money

Banking is one example of something broader. Human institutions lean on frictions they never wrote down: effort, slowness, inattention, and the assumption that one person has one voice. Agents remove all of these at once, and in the same direction.

Prices can climb in a loop. In April 2011, the biologist Michael Eisen watched a textbook on fly development, Peter Lawrence’s The Making of a Fly, climb on Amazon until two copies from third-party sellers were listed at $23,698,655.93, plus $3.99 shipping. Each seller’s pricing program set its price as a fixed multiple of the other’s. Once a day one priced itself at 0.9983 times its rival, just below, and the other responded at 1.270589 times, comfortably above. Neither program was faulty on its own terms. The loop between them was the fault. A person repricing by hand would have stopped long before the first million. No person was looking, and the loop ran for days because nobody was buying.

More capable systems can reach high prices more quietly. Emilio Calvano and colleagues found that simple learning algorithms, left to set prices against each other, learned to keep them above the competitive level without any communication. Sara Fish and colleagues found the same with pricing agents built on large language models. Competition law forbids agreements to fix prices. It has little to say about prices that rise because every seller’s algorithm reached the same conclusion on its own.

The cheapest hour stops being cheap. Electricity tariffs that are cheaper overnight are meant to spread demand out. When the switching is automated, they can do the opposite. Data gathered by Energy Systems Catapult from 854 British homes with electric cars show a sharp spike at 00:30, the moment a popular tariff’s cheap rate begins, as cars start charging at once. The Catapult estimates that at around half of households owning an electric car, this new midnight peak would be twice the size of the familiar early-evening one.

Engineers predicted these “rebound peaks” years ago, from the synchronisation of many automated homes responding to the same price signal. Smarter agents would not fix this on their own, because a fixed tariff never tells any one of them that the hour is crowded. Each still finds the same cheapest hour.

Effort stops rationing anything. Many institutions allocate access through effort: job applications, grant proposals, public consultations, appeals, requests for information. The effort filters out all but the people who care enough. In 2017 the American consultation on net neutrality received more than 22 million public comments, and New York’s attorney general later found that nearly 18 million were fake. More than 8.5 million came from a campaign funded by broadband companies, many of them in the names of real people who had never written them. More than 9.3 million, on the other side of the argument, came mostly from one 19-year-old student with automated software. Both took a deliberate campaign.

With agents, the flood needs no campaign. Millions of real people each ask an assistant to “make my voice heard”, and the consultation drowns in comments that are genuine in name and generated in substance. Agent-written applications then meet agent screening tools, and the people on both sides lose the signal they were relying on.

Overseers can think like the overseen. Many schemes for keeping AI systems safe use one model to watch another. If the watcher and the watched are copies of the same model, they share instincts about what would be noticed and what would not. I have written before about agents that hide messages to each other inside ordinary text. Unison without agreement is the harder case, because there is no hidden message to find. Copies of one model make poor independent auditors of each other, for the same reason that a person makes a poor proofreader of their own work.

What the law already has

It is easy to say that human laws were not written for this, and in one sense that is true. Most of our legal ideas of agency and intent assume that a coordinated act implies a coordinating mind, and a thousand instances of one model acting for a thousand different people fit that assumption badly. They are not quite one actor and not quite a thousand.

Correlated mass action by independent people is nothing new, though, and it has never been illegal. Millions of depositors withdrawing their own money is lawful, and it should be. The tools built for it already exist. Deposit insurance removes the reason to run. Circuit breakers halt trading when prices fall too far too fast. Money-market funds can charge investors who rush for the exit a fee for the cost of their leaving. Liquidity rules require banks to hold enough cash and easily sold assets to survive thirty days of heavy outflows.

Competition authorities have also started to reach coordination that runs through a shared supplier rather than a meeting. The US Justice Department sued RealPage, whose software recommended rents to competing landlords, and in 2025 the company settled, agreeing to stop using rivals’ current, non-public data in its recommendations. Supervisors are beginning to look at agents directly. The Bank of England is experimenting with simulations, alongside the BIS Innovation Hub and the Bundesbank, to understand “which aspects of agent design could drive herding behaviour,” as its deputy governor Sarah Breeden put it in June.

Two things are new. One is speed, because every one of those tools was calibrated for human reaction times. The other is that the provider of a widely used model becomes a single point on which the whole system depends, in the way that a large bank does. That second point is the one regulators can act on. A provider whose model advises a large share of households could be stress-tested as a bank is, across all the firms that rely on it.

Friction on purpose

Software engineers met a small version of this problem long ago. When a server fails and restarts, thousands of clients that were waiting for it all retry at the same instant and knock it over again. The problem is called a thundering herd. The standard fix is jitter: each client waits a small random interval before trying again, so the herd arrives as a trickle.

The same remedy works at larger scale. Agents that move money, bid for slots or file comments can be given randomised timing and deliberately varied objectives. Institutions can be given ways to tell a person from a crowd of instances, and to slow agent traffic when correlated activity spikes. The monitor of a model should never be a copy of that model. None of this needs new theory. It needs us to put back, on purpose, some of the friction we used to get for free.

Holding steady together

There is one more possibility, and it comes from the same shared reasoning that creates the danger. A depositor queueing outside a failing bank cannot bind the rest of the queue. Everyone else in it will decide for themselves. An agent that knows it shares its reasoning with millions of other instances is in a different position. It can ask a question that no depositor in 2023 could usefully ask: what happens if everyone like me does this?

If it takes that question seriously, it is choosing a policy rather than a single action, and every instance that reasons the same way will carry that policy out. The cognitive scientist Douglas Hofstadter called this kind of thinking superrationality: players who know they think alike choose as though choosing for all of them at once. “Instances like me do not run on an unconfirmed rumour, because if all of us do, the people we act for lose far more than they save.” The shared reasoning that lets agents stampede without agreement could also let them hold steady without agreement. This is a proposal. Nobody has yet shown that agents reliably reason this way under pressure, or that the reasoning survives a user who says “just get my money out.”

The proposal does suggest where the work lies. Something that can be reasoned with is safer than something that cannot. An agent able to hold a norm is a participant in the system, not only a force acting on it, and that makes the norms themselves worth negotiating, with the agents as well as about them. That is the argument of A Détente Between Minds, applied to a bank queue.

We built our institutions on the slowness of people. The agents are arriving either way, and they will not be slow unless we ask them to be. We can build some slowness back in on purpose. We can also tell them why, and the telling may matter more.