The Ledger That Keeps Something
Every vending machine has a door at the back. You never see it opened, which is rather the point, but somebody holds the key. They come round once a week, or once a fortnight, empty the coins, refill the spirals, clear the can that jammed on Tuesday, and, when the supplier puts its prices up, change the number on the button. The machine carries out its terms exactly as written, the same way for everyone who walks up to it. The person with the key decides what the terms are.
I ended The Ledger That Does Something on that door, and I want to start from it, because the two posts before this one were both about removing the person standing in front of the machine, and neither managed to remove the one standing behind it. The Ledger That Costs Something showed how Bitcoin took the keeper out of a ledger of payments by making every line cost energy anyone can check. The Ethereum post showed how a ledger could run agreements as well as payments, with code standing where the clerk used to. Both are real achievements. Both are also, as I said at the end of each, prototypes, and the most useful thing a prototype produces is the list of what it got wrong.
This post is that list, and it's also the reason the series exists. Back in July I asked Claude a question I'd been turning over for a while. If you were an AI and needed to buy things, pay for services, or hire a person to do a job in the physical world you couldn't do yourself, which cryptocurrency would you use? It looked over the top twenty and answered USDC, a dollar stablecoin, running on Ethereum and the cheaper rollups built on top of it, with Ethereum itself as its pick if it had to hold something that wasn't pegged. Its reasons were sensible. A contractor doesn't want to be paid in something that might be down a fifth by the time they cash it out. The payment needs to be programmable, so the money can sit in escrow until the job is done. The fees have to be small enough for small jobs, and whoever gets paid needs to be able to turn it into something useful. Bitcoin, it said, was the purist's answer, and a bad fit for paying somebody reliably for a day's work.
I've thought about that answer more than I expected to. Read it with the last two posts in mind and it's a tidy summary of everything they found. The most capable non-human buyer I could ask, choosing freely between every major coin, picked a company's promise to pay dollars, with a freeze function in its contract, recorded on a ledger with no keeper. It reached straight past both proofs of concept for the keeper sitting inside them. I think it was an accurate reading of where these systems actually are, and that's the uncomfortable part.
So the questions I've been asking since July are the ones this post and the two after it try to answer. Why do keepers keep coming back? What would a ledger have to look like for a machine to hold and spend money on its own, with no person in the middle? And if the keeper can't be removed, which I've come to believe, what kind of keeper should it be?
There are no new mechanics here. Everything technical this post leans on was laid out in the Bitcoin and Ethereum posts, and where I lean on it I'll link back rather than explain it twice. We'll start with a short history of keepers, from the first ledgers through banking and computing to the two proofs of concept. Then the three ways the keeper got back in, what the wrappers wash on the way through, and what time will take regardless. Then the customer neither system was designed for, the possibility that the next keeper is software, and why, with all of that on the table, we still need both chains. It ends with the questions I think whatever comes next has to answer.
A Short History of Keepers
The first keepers were priests. Some of the oldest writing we have is accounting: clay tablets from Mesopotamian temples recording grain, livestock and labour owed and delivered. The temple held the stores, wrote the lines, and its authority was what made a line worth anything. Coins came later and moved the keeper without removing it. A coin carries its keeper's face. The stamp is a promise from whoever minted it about weight and purity, and debasement, quietly shaving silver out of the coinage, is about as old as coinage itself.
Banking moved the keeper again, from the mint to the counting house. A goldsmith's receipt for gold in his vault could be passed from hand to hand, and before long the receipts were doing the work of the gold. The goldsmith then noticed he could write more receipts than he had gold, so long as everybody didn't come back on the same morning. That's the fractional reserve I described in The Debt Beneath the Debt, and it's where the keeper stopped merely recording and started issuing. Double-entry bookkeeping, set down in print by Luca Pacioli in 1494, made one keeper's books checkable by another, which is the beginning of the auditor, the regulator, and every other arrangement for watching the keeper I listed at the start of the Bitcoin post.
Computing automated almost all of it. Banks were among the first large commercial buyers of computers, and over the second half of the twentieth century the ledger moved from a clerk's hand to a mainframe, and from there to a database most bankers never see. In places the software is older than the bankers. A surprising share of the world's banking still runs on COBOL, a language first specified in 1959, maintained by a dwindling number of people who learned it before most of their colleagues were born (if your card payment has ever cleared on a Sunday night, there's a fair chance it was approved by code in a language older than the moon landing). Your balance is a row. Moving money is an update to that row and a matching update to somebody else's. The work of the keeper was handed to software, and the keeper itself stayed exactly where it was, because somebody wrote the software, somebody runs the database, and somebody decides which updates are allowed. The computer made the keeper faster and cheaper. It did nothing at all about the keeper's interests.
Then came the two proofs of concept, and for the first time the keeper was taken out of the software as well. Bitcoin put the ledger on thousands of machines that don't trust each other and replaced the keeper's judgement with a rule every node can check: the chain with the most work is the ledger. Ethereum extended that to agreements, so an escrow or an exchange or a loan could run on the same terms for everyone, with nobody at the desk. Inside the software, both worked, and both still work. Nobody has ever forged a Bitcoin signature, and the only block that broke Bitcoin's rules on the main chain, the 2010 bug that minted 184 billion coins, was rewound within hours.
The trouble is everything outside the software. A ledger records claims on things people want, kept by people with interests, inside countries with laws, and as soon as either chain was useful enough to matter, the world it was built to route around came back in. The keeper was removed from the code. It was never removed from the politics, and the rest of this post is about the routes it took back.
Where the Keeper Got Back In
Go back through the last two posts and count the keepers. I did, and there are well over a dozen, but they came back through only three doors.
The Edge
A ledger can only be keeperless about what lives entirely inside it. Bitcoin's ledger records bitcoin, and bitcoin exists nowhere else, so the only keeper it needs is the rule. The moment a line points at something outside the ledger (a dollar in a bank, a bond with a custodian, a plot of land, a bitcoin in a vault standing behind a token on another chain) the line inherits that thing's keeper. The ledger can record the claim perfectly. It has no power at all over whoever holds the thing being claimed.
That's where most of the money went. The $180 billion or so of stablecoins on Ethereum are dollars in bank and custody accounts, represented on-chain by a company's promise with a freeze function in the contract. Wrapped bitcoin is bitcoin held by BitGo or Coinbase. Tokenised Treasuries are Treasuries at a custodian. Around 80% of the bitcoin in US spot ETFs sits with Coinbase, and the investor never touches a chain at all. Even the land registry I sketched in Land on the Ledger needed a statute to make the token the deed and a court-signed override to freeze or reassign it, because a plot of land doesn't care what a contract says. Every bridge between a keeperless ledger and the world has a keeper standing on it, and so far the bridges carry most of the traffic.
The Rules
The second door is the one The Short Circuit was about. A ledger's rules are only as fixed as the agreement of the people running its software, and every time something went badly enough wrong, that agreement was asked to decide. In 2010 it rewound a block that created 184 billion bitcoin. In 2016 it reversed The DAO, where the code had done exactly what it said and the community disliked the result. Ethereum's monetary policy has been rewritten at least four times. BIP-361 would freeze coins whose owners can't or won't migrate them before a quantum computer arrives, and Vitalik Buterin has already published an emergency plan to roll Ethereum back if one does.
Each of those was argued in public, by people acting in good faith, and I'd probably have supported most of them. That's what makes them instructive. Nobody broke a rule in any of them. The rules were moved by the people in a position to move them, which on a ledger with no keeper turns out to be whoever writes the client software and whoever chooses to run it. In The Short Circuit I said you can't find out whether a ruler has shrunk by laying it against itself. A blockchain's rules are measured against the blockchain's own community, and when that community decides the ruler should be shorter, it is.
Scale
The third door has nothing to do with law or politics. It's efficiency. Anything cheaper to do in bulk ends up done by a few, and almost everything in both systems is cheaper in bulk. Mining smooths out its lottery through pools, and a handful of pools produce most of Bitcoin's blocks. Staking is easier through a provider, so a fifth of all staked ETH sits with Lido and much of the rest with exchanges and custodians, which is where I put mine. Building profitable blocks takes capital and specialist software, so three builders produced roughly 90% of Ethereum's blocks across two years, and by this summer one was producing around half. Rollups were cheaper to launch with a single sequencer, and most still have one. Even the software concentrates: about half of Ethereum's execution clients run Geth.
Nobody planned any of it. Each step was the reasonable choice for the person taking it, which is how the keeper comes back when nobody is trying to bring it back. And when the Tornado Cash sanctions arrived in 2022, the decision about what most of Ethereum's blocks would carry fell to a few builders and relays. The thousands of validators signing those blocks were never asked.
What the Wrapper Washes
In Much Ado About Laundering I described war as a laundering machine. Money goes in through fear and urgency and comes out the other side clean, along with the legitimacy of the people who handled it. The edge of a keeperless ledger works the same way, with trust in place of money.
Follow a stablecoin dollar through it. It starts as a deposit in a bank account belonging to the issuer, a keeper's liability in the most ordinary sense. It's minted onto Ethereum as a token. From then on it moves through contracts with no keeper, settles in twelve seconds, passes through exchanges and lending pools and escrows, and everywhere it goes it picks up the reputation of the ledger it's travelling on: transparent, programmable, open to anyone. What it never loses is the keeper at the start. The bank account is still there, the freeze function is still in the contract, and the issuer still answers to a regulator and a court. The token has been washed of the look of a keeper while keeping the keeper.
My BETH was the same thing on a smaller scale. It was a promise from Binance, made liquid, traded by people who never saw the ETH behind it, and priced as though it carried the properties of the chain the ETH was locked on. It carried Binance's.
The washing runs the other way too. The spot ETFs took the one asset whose whole point was that it needed no custodian, put it in a custodian, and sold the result to pension funds and brokerage accounts through the same plumbing as everything else on Wall Street. That's the existing system borrowing Bitcoin's legitimacy (the seventeen years, the hashrate, the energy anchor) and wrapping it in a keeper of its own. The fund holder gets the price of bitcoin and none of its properties, and the keeper gets to point at the most secure ledger on Earth while holding the key.
I don't want to overstate this. Wrapping is useful, and most of the people doing it are doing what they say. The pattern from the laundering essay still holds: something passes through a process that makes it look clean, and the cleaning is cosmetic. The chain checks every signature and every line. It has no way of checking the promise behind a token, and the further a token travels from its issuer, the easier that promise is to forget.
The chains have laundered money in the ordinary sense as well. Tornado Cash was sanctioned because it had been used to move stolen funds, including, according to the US Treasury, hundreds of millions of dollars taken by North Korea's Lazarus Group. That's the other face of a ledger with no keeper: nobody there to refuse the transaction. The response was to invite the keeper back in through the builders and relays, which settled whether the keeper would return by deciding who it would be.
What Time Will Take
Some of the keeper's routes back in are open now. Others will open on a schedule whether anyone chooses them or not, and both earlier posts ended their technical halves on them.
Bitcoin's security budget shrinks every four years. The subsidy halves, and unless fees rise to replace it or the price doubles each time, the energy defending the chain falls with it. Nobody yet knows whether the fee market will be deep enough once the subsidy is negligible, and the problem has a deadline written into the code.
Both chains' signatures rest on mathematics a large enough quantum computer would break. A third of all bitcoin, and somewhere around 55 to 60% of all ETH, sits behind public keys already visible on the ledger. Ethereum is better placed to respond because it can change almost anything, and Bitcoin is worse placed for the same reason in reverse.
Bitcoin's proof of work is welded to SHA-256, and an industry of single-purpose machines exists to compute it. If the hash ever has to change, the hard fork that changes it bricks every one of those machines on the day it activates.
Ethereum's state, every account and every byte every contract has ever stored, only grows, and keeping it manageable is one of the oldest unsolved problems on its roadmap. A node that costs too much to run is a node fewer people run.
Look at what each of those needs to resolve it. None can be fixed by the protocol as it stands. Every one requires a decision: a new fee policy, a migration deadline, a freeze, a new hash function, a change to what a node has to store. A decision needs someone to make it, and on these ledgers that's the same small group of developers and the same agreement among node operators that reversed The DAO. Time wears a ledger down, and it also sends it back to the pen, again and again, and every crisis that needs a decision is a moment when the keeper is, briefly and legitimately, in charge.
That's the part I find hardest to argue around. Even a perfectly designed keeperless ledger would eventually meet a problem its designers didn't foresee, and the only mechanism either chain has for that is people agreeing to change it.
The Customer Nobody Planned For
Both systems were designed with a particular user in mind, and that user was a person. Satoshi's paper is about electronic cash between two parties who'd rather not go through a bank. Ethereum's account model assumes someone holding a key in a wallet, deciding what to sign. Even the keepers that came back are built for people: an exchange account needs a passport, a bank account needs an address and a credit history, a card needs a name on it.
The customer arriving now has none of those things. AI agents already book, buy and call paid services on behalf of the people who deploy them, and the more capable they get, the more of that they'll do without anyone approving each step. An agent that needs to pay for an API call, rent an hour of compute, buy a dataset or, as my July question put it, hire a person to do something in the physical world it can't do itself, needs four things from whatever it pays with.
It needs to hold funds. That means a balance of its own to draw on, rather than a person's card or a login to somebody else's account. No bank will open an account for a piece of software, and no regulator I know of has a category for one.
It needs to spend within limits someone else has set: a budget, a list of counterparties it may pay, a ceiling per task, a date after which the permission lapses. That's what account abstraction, which Ethereum built into ordinary accounts in 2025, makes possible. An account can carry its own rules, so an owner can hand an agent a key that can only do what the owner allowed.
It needs to pay at machine scale. Agents don't make a handful of large payments a month. They make thousands of tiny ones, fractions of a cent for a query or a second of compute, and card networks were never priced for that. The x402 standard, which gained the backing of Google, OpenAI and Circle this April, exists to let an agent pay for a service in stablecoins in the middle of an ordinary web request.
And it needs to get paid and settle without a person countersigning every step, because the whole point of handing a task to an agent is that you aren't standing over it.
Both chains can, technically, do all four, and Ethereum already does most of them in production. That's the strongest argument for them anywhere in this series. They're the only financial rails in existence that accept a non-human participant on exactly the same terms as a human one. A key is a key, and the network doesn't ask who's holding it.
Now look at how it's actually being done. The agent spends USDC, which has a freeze function. It does so on a rollup with a single sequencer, which can refuse or delay its transaction. The key it spends with sits in a wallet set up and controlled by the company that deployed it, which can revoke it at will. The agent has money. It doesn't have its own keys, and the keepers between it and its money are exactly the ones the last two posts found everywhere else.
Agents as Keepers
Which brings me to the possibility I think matters most, and the one I'll leave the next two posts to work through.
So far the keeper has come back as institutions: custodians, pools, builders, issuers, developers. The next time it comes back, I think it comes back as software. Agents will spend money, and before long they'll also hold it on behalf of others, escrow it between parties, route it, invest it, and decide when to release it, which are the jobs keepers have always done. Some of that is already happening. The searchers and builders who decide the order of Ethereum's transactions are, for the most part, automated systems running strategies nobody reviews block by block, and MEV is what a keeper looks like when it's a program. The vending machine from the last post is growing hands.
There's a real argument that an agent makes a better keeper than a person. It doesn't get tired. It can't be bribed in any ordinary sense, because it wants nothing. It doesn't steer a contract towards its cousin. Given a clear rule, it applies the rule the same way at three in the morning as at noon, for everyone who walks up to it, which is the vending-machine promise extended to work that needs judgement.
The catch is the one The DAO taught us ten years ago. An agent does exactly what it's been told, and somebody did the telling. Whoever writes its instructions, chooses its objective and controls its updates holds the pen, and everything I said in The Short Circuit about the floor that writes the standard applies to them. The agent can't be greedy. Its owner can, and the agent will carry out the owner's greed faithfully, at scale, without complaint. It can also be told things by people who aren't its owner, through instructions slipped into the data it reads, which is a way of moving the pen nobody has fully solved.
Then there's the question of what you'd check. Bitcoin asks you to check the cost, and anyone with a laptop can. Ethereum asks you to check the code, which anyone can read. An agent's behaviour comes out of a model nobody can read line by line, plus instructions that may or may not be public, plus whatever it has seen since it was deployed. A keeper you can't inspect is a keeper you're trusting, which is where this whole story started.
And there's a choice nobody gets to avoid. Seal the agent, so its instructions can't be changed after deployment, and you've built something that can't be corrupted and can't be fixed, which is Bitcoin's problem with its hash function. Leave it upgradable and you can fix it, and someone holds the key to the upgrade, which is Ethereum's problem with everything.
That choice gets much sharper when you ask what agents might eventually keep. Registers of land and intellectual property are records, and ledgers are already good at records. Tax collection, pensions and benefits are transfers, rules applied to balances, and those can be automated. Policy, the setting of rates and rules and who qualifies, is the pen itself. Courts and councils and the enforcement behind them are judgement and force, which is the edge in its rawest form. The closer a function sits to record-keeping, the more of it a ledger could take over. The closer it sits to judgement, the more certain it is that a keeper remains, and the question becomes what kind. An agent that runs a pension fund for forty years without ever speaking to a person is either the most trustworthy keeper ever built or the most dangerous one, and which it is depends entirely on who could change it, how, and whether anyone could check.
Why We Need Them Anyway
After two posts and most of a third on what's wrong with them, I owe an account of why I still think Bitcoin and Ethereum are the only two in this space worth taking seriously, and why we need them.
The first reason is the simplest. They're the only working proofs we have. Before 2009 a ledger without a keeper was a thought experiment, and before 2015 an agreement without one was a vending machine. Both now run, carrying real value, on networks nobody owns, and both have survived crashes, bans, bugs, hacks, forks and a decade of people trying to break them. Every flaw in this post was found because they ran. You can't learn what goes wrong with a design nobody has built.
The second is the one I came to through the July question. They're the only financial rails that let a machine take part on equal terms. Everything else checks for a person at the endpoint, and the world is about to fill up with endpoints that aren't people. Whatever eventually replaces these systems, the reason anyone will know how to build it is that these two proved a key could be enough.
The third is the one I think gets overlooked, and it's what I'd point to if someone told me the returning keepers made the whole exercise pointless. Every keeper I've listed in this post, I could list because it's visible. I can tell you that Lido holds a fifth of staked ETH, that one builder produced half of this summer's blocks, how much ETH is burned each day and how much sits behind exposed keys, because the ledger records all of it and anybody can count. The keepers in the existing system are mostly invisible from outside. Nobody outside a bank can audit what stands behind its deposits in real time, and the inquiries into keepers are usually run by other keepers.
In The Short Circuit I said oversight fails at the top because you can't measure a ruler against itself. These ledgers are the first ruler that sits outside the institutions it measures. They haven't removed the keeper, and I've spent most of this post showing they can't. What they've done is make the keeper countable. A keeper you can see is a keeper you can do something about, and the third check I described in that essay, the people closing the loop, depends on being able to see where the loop is broken.
So we need them, flaws and all: because they're the proof, because they're the only door open to the new customer, and because they're the only place the keeper has to work in public.
What the Next Ledger Has to Keep
At the end of the Bitcoin post I said its flaws were the specification for whatever comes next, and at the end of the Ethereum post that Ethereum's were a long addition to the same document. Put together, the specification turns out to be a short list of questions rather than a list of features, and they're the ones I've been carrying since July.
How do you tie a currency to energy, even as a proxy, without rebuilding the seventeen years of infrastructure Bitcoin has already paid for? The energy anchor is the one property no keeper can suspend, and nobody is going to rebuild a zettahash from scratch.
How do you get the security of proof of work at the energy cost of proof of stake? Work defends deep history and grows more expensive to rewrite the older it gets. Stake defends the tip cheaply and needs someone to tell a new node where the chain is. Each is strong exactly where the other is weak. And if the work is going to be done anyway, can it be work that produces something besides heat?
How do machines hold and spend money with other machines, with no person in the middle, on rails with no freeze button and no single sequencer standing between an agent and its funds?
And how do you build something that could one day carry the functions of governments and institutions, kept by agents that never deal with a person once they're deployed, without handing the pen to whoever deployed them? If the keeper can't be removed, and seventeen years of evidence say it can't, then it has to be replaced with one greed can't reach. That means a keeper with no interest of its own, whose interest no single party can set, whose behaviour anyone can check, and which can still be corrected without being captured by whoever does the correcting. Nobody has built one. I'm not sure anybody has properly described one.
The next post takes the two ledgers as they are and asks what the world would look like if we built on both, each doing the job it does best. The one after that asks what it would take to replace them.
The door at the back of the vending machine is still there. For most of history someone came round with the key, and every improvement we made changed who that was, or how fast they worked, or how hard they were to watch. The two proofs of concept locked the door and threw the key away, and within a few years people had cut new ones, because a machine with nobody behind it can't be restocked when the world changes. The next machine may well restock itself. It'll order from the supplier, pay the invoice and clear the jam with nobody standing at the back. Somebody will still have written what it does when the supplier puts its prices up, and the question for everything that comes next is who.