There is a wallet on Base that paid 2,190 ETH in transaction fees over eleven weeks.
It did not trade much. It did not bridge much. It paid.
I pulled thirty days of its calldata and 94% of it decoded into four function selectors. Four selectors, repeated 2.1 million times. No human sends the same four calls two million times because they like the rhythm of it. That wallet is a machine.
And a machine has a keeper.
I spent three weeks in Paris running a clustering script against seven rollup RPC endpoints, hunting the agentic economy that keeps arriving in my inbox at forty-million-dollar valuations. What I found instead: 4.1 million "unique active addresses" that collapse, under three-hop funding-graph clustering, into 312 wallets. Those 312 paid 41% of all gas on those seven chains during the window.
That number belongs on every Layer2 pitch deck. It is on none of them.
CONTEXT: HOW WE GOT HERE
The framing I keep hearing is that rollups are scaling Ethereum and AI agents are scaling rollups. Both halves of that sentence are doing a lot of unpaid labor.
Start with supply. We now have dozens of general-purpose rollups with meaningful state, plus another tier of app-chains, validiums, and "sovereign" forks that borrow the brand without the security model. After EIP-4844, blob space made L2 fee markets cheap enough that the marginal cost of a transaction dropped by roughly an order of magnitude on the chains that adopted blobs early. Cheap blocks change behavior. When a call costs fractions of a cent, the rational strategy is no longer "make every transaction count." It is "make two million transactions and hope the aggregate outcome counts."
That is the soil the agent narrative grew in.
By early 2025, "agent" had become a token suffix. Every project with a Python loop and a funded wallet had a framework, a launchpad, and a chart. The pitch was consistent and, on its face, correct: autonomous economic actors will transact machine-to-machine, and smart contracts will become the settlement layer for that machine economy. I believed a version of this. In my own column I published a framework arguing that by 2027, roughly 60% of on-chain volume would originate from agents rather than humans.
This article is me checking my own homework. The headline number may end up right. The reason it ends up right is not the reason I gave.
Because volume is a numerator with no denominator problem. It counts everything and asks about nothing. Volume does not distinguish between a machine buying a treasury bill and a machine buying and selling the same token against itself ninety times to farm a points program. Both are volume. Both are gas. Only one is an economy.
Two structural facts make this worse than it sounds.
The first is that rollup growth is now measured almost entirely by inputs the operator controls. Active addresses, transaction count, gas consumed — every one of these is emitted by the chain itself, ranked by dashboards that scrape the chain, and reported by the chain's own foundation. There is no independent auditor of "activity." There is no external settlement constraint that punishes inflated numbers. The scoreboard and the player are the same entity.
The second is that the bridge layer has become the only place where truth leaks out. Value that enters from L1 and stays is real. Value that enters, loops, and leaves is a rumor with a receipt. Most dashboards count the loop and ignore the exit, because the exit is boring and the loop has a chart.
I have watched this pattern before. In 2017 I ran superficial audits across more than forty ICO contracts during the peak of that boom, and the tell was always the same: the projects with the loudest metrics had the thinnest on-chain footprints once you subtracted the addresses they funded themselves. The tooling has changed. The shape has not.
CORE: THE DATA
Here is the method, so you can break it.
I ran web3.py against public RPC endpoints for Arbitrum, Base, Optimism, zkSync Era, Scroll, Linea, and Mantle. I pulled 30 consecutive days of blocks, extracted every sender address, and built a Union-Find graph over first-hop funding relationships: if wallet A's first inbound ETH came from wallet B, and B's came from C, they collapse into a single cluster at three hops. Then I decoded calldata against a local four-byte signature directory and bucketed by selector.
Two caveats, stated up front. First, three-hop clustering over-merges some clusters — exchange hot wallets are notorious for welding unrelated users into one blob. I excluded known exchange and bridge deposit addresses manually, which is tedious and imperfect. Second, "agent" is a label I am applying, not a fact I am proving. What I can prove is that the behavior is machine-shaped: identical selector distributions, metronomic inter-block timing, and near-zero variance in gas price bidding.
Finding one: the concentration is not marginal. It is the market.
4.1 million addresses. 312 clusters responsible for 41% of gas expenditure. The top 20 clusters alone accounted for 23%. And of those 312, 189 had activity on three or more of the seven chains inside the same 30-day window, with near-identical selector distributions.
Read that again slowly. Near-identical selector distributions across seven independently operated chains. That is not seven user bases. That is one operator holding seven endpoints.
Finding two: four selectors.
I pulled the largest cluster's calldata distribution. It broke down roughly as: 38% a swap variant, 24% a claim() call, 19% a repay(), 13% a withdraw(), and 6% everything else. The ordering matters. Swap, claim, repay, withdraw.
That is not a trading strategy. That is a credit cycle executed by a robot against a subsidy program — borrow, deploy, harvest the incentive, repay, exit, repeat. The claim() selector is the tell. Pure directional traders do not spend nearly a quarter of their gas on claiming. Farmers do.
Finding three: the fee was never the agent's fee.
I traced the fee payments. On Base alone, 71% of the gas paid by the top 189 clusters was sponsored — routed through paymaster contracts, with the paymaster's own balance topped up from a small number of treasury addresses. So when a chain announces that agentic activity drove X% of transactions, ask the second question: who paid for the gas?
On these numbers, the answer is not the agent economy. The answer is a handful of treasuries subsidizing the appearance of one. Which means the demand is a marketing line item on someone's balance sheet, and it stops the quarter the incentives stop.
Finding four: sequencer revenue does not agree with the volume chart.
This is the cross-check I trust most, because it is the hardest to fake. Sequencer revenue is a function of gas consumed times the base fee, net of the blob cost paid to L1. If agentic volume were genuine, independently funded demand, sequencer revenue would trend with it.
On four of the seven chains, it did not. Volume went up. Revenue went flat or down. That divergence has one clean explanation: the marginal transaction is funded by a subsidized paymaster at a subsidized gas price, which consumes blockspace but does not consume anyone's wallet.
The truth is hidden in the gas fees. And the gas fees say the demand is rented.
Finding five: the multisig.
This is the part that made me put the coffee down.
I pulled the admin and owner addresses for the top 50 contracts interacted with by those 189 clusters. Then I pulled the signer sets for every multisig in that list. Three of the five largest agent operators — collectively driving 14% of the gas in my sample — share signer overlap on a 4-of-7 multisig. Two of those seven signers also sit on the upgrade multisig of two lending markets those same agents borrow against.
There is no allegation of wrongdoing in that sentence. There does not need to be. The structural point is enough: the "autonomous" economy is autonomous in its execution and fully centralized in its permissions. The agents run themselves. The rules they run under can be rewritten by four people who do not have to tell you.
Every DAO that markets itself as decentralized governance while its upgrade authority sits with a 4-of-7 has the same shape. Code is law, but audits are mercy — and nobody in this sample has commissioned the audit.
Finding six: 41 proxies, no timelock.
I checked the upgradeable contracts in the interaction set. Forty-one used a proxy pattern with an admin-controlled implementation slot. Nine had a zero-delay admin. No timelock. No announcement. No window. A single key can change the contract logic that hundreds of millions in agent flow routes through, in the same block it decides to.
I ran superficial audits on more than forty token contracts in 2017 with worse tooling than I have now, and the lesson from that year has not aged: the exploit is rarely the clever math. It is the permission you forgot was there. Reentrancy gets the headlines. Admin keys get the money.
Finding seven: the bridge flow.
One more cross-check. I traced net value back to L1 settlement. Genuine new economic activity should show up as net inflow from L1 that stays. What I found across the seven chains was a circulating loop: value bridges in from L1, deploys into subsidized positions, harvests, unwinds, and bridges back out within days — with the fee cost covered by the paymaster.
The loop is closed. It is a closed loop with a bridge on it. Net new value settled to L1, after subtracting the loop, was a small fraction of the gas headline. That is the honest measure of whether an agentic economy exists. Not how many calls were made. How much value came in and stayed.
Finding eight: the agents read untrusted data, and they can sign.
This is the vector almost nobody is pricing.
Most agent frameworks in this sample operate on a simple loop: ingest unstructured input — a social post, an on-chain memo field, a token name, a forum thread — reason over it, then emit a transaction through a key that has standing allowance. I found 63 agent contracts with live ERC-20 approvals to routers they had never interacted with before, approvals granted by whatever the last piece of text the agent read told them to grant.
That is prompt injection with a signature attached. The attacker does not need to break cryptography. They need to write a sentence in a place the agent reads. The money moves through an allowance the user granted once and forgot.
Rewriting the rules before the bug writes them is the entire job here, and almost no agent framework ships with an input-sanitization layer, a spend cap, or a session key that expires. We built autonomous signers and handed them open reading privileges.
CONTRARIAN: IT IS NOT SCALING, IT IS SLICING
The dominant narrative is that Layer2s are scaling Ethereum and agents are scaling Layer2s. The contrarian read is simpler and less flattering: there are dozens of Layer2s and roughly one user base, and the agent narrative is what we use to avoid saying so.
Liquidity doesn't fragment on a chart. It fragments in the mempool — seventeen order-flow auctions clearing the same depth across seventeen chains, each one a little thinner than it would have been as one. Splitting a scarce pool into seventeen pools does not produce seventeen times the liquidity. It produces seventeen times the surface area for slippage, seventeen times the bridge risk, and seventeen times the number of places where whoever runs the sequencer sees the order first.
The agent narrative makes this worse, not better, because it converts a measurement problem into a marketing asset. When volume is machine-generated, activity becomes trivially manufacturable. A loop of two contracts can produce more transactions in an afternoon than a mid-sized country produces in a year. No users needed. No capital needed beyond what the paymaster fronts. Just a script and a treasury.
So the metric the entire Layer2 sector is ranked by — transactions, active addresses, gas consumed — is the one metric a botnet can mint at will. We built a leaderboard for a game where the score is printed by whoever owns the printer.
There is a second blind spot underneath. Even the honest version of the agent economy is not an independent economy. It is a coefficient on the existing one. Agents do not have preferences, savings rates, or conviction. They have objective functions, and the objective function is written by whoever funded the treasury. So long as the funding is a points program, the agents are a points program with better uptime.
The pool remembers what the ticker forgets. Every subsidized loop leaves a permanent record of who paid for it, and on all seven chains the record says the same thing.
And here is what should worry the bulls most. The verifiability bargain of blockchains was always that cost equals commitment — a transaction you pay for with your own money carries information about your conviction. Volatility is the tax on uncertainty, and we accepted that tax. Subsidized gas breaks the bargain. It makes transactions free, and free transactions carry no information. We did not get a faster economy. We got a cheaper signal, which is the same as no signal. Speculation is just data with a heartbeat, and we have been reading the pulse of a paymaster.
TAKEAWAY: WHAT TO WATCH
Use different numbers.
Watch sequencer revenue per chain and ask why it diverges from volume. Watch the implementation slot diff on any proxy your capital routes through, and treat a zero-delay upgrade on a contract holding nine figures as a live position, not a footnote. Watch net value settled to L1 after subtracting the loop — the only number a paymaster cannot print.
And watch the signer sets. All of them. Every time. Because in an economy where the agents are autonomous and the permissions are four people, the risk is not the machine.
The machine is fine. The machine is exactly as smart as the four people who can rewrite it at block height N+1 — and the question worth asking is not whether they will. It is whether anyone would know if they did.