The Mercenary Machine: How the AI-Agent Economy Is Quietly Starving Every Layer 2

On the morning of March 4, at 03:14 UTC, a wallet I had been tracking for eleven weeks executed 4,118 transactions inside a single hour. The largest of them moved $0.0031. None touched a governance contract. None held the rollup's token. None paid a discretionary fee, because a paymaster had abstracted gas into a rounding error the way a cloud provider absorbs the cost of a heartbeat ping. The agent was buying nothing you would recognize as an asset. It was paying for inference, verifying a signature, settling a micro-debt between two machines that will never meet.

By the end of that week the rollup had printed the highest transaction count in its history and its token had fallen 6.2%. Its governance forum filled with people asking the same question in different fonts: if this is adoption, why does it feel like a leak?

I have spent twenty-six years watching this industry mint a narrative, fall in love with it, and then discover that the thing it built does not quite match the story it told. This particular gap is wider than most, because the newest and fastest-growing class of user — the autonomous software agent — is also the most indifferent. It does not tweet. It does not farm points with sybil wallets. It does not care about your airdrop. It has a stablecoin balance, a latency budget, and an objective function, and it is quietly renegotiating the economics of every chain it touches.

Tracing the ghost in the machine has always been my job. This ghost is quieter than the ones I grew up with. That is precisely what makes it dangerous.

Context: the cycles that taught us to mistake plumbing for demand

Every market cycle in crypto has a founding myth, and every founding myth eventually collides with the plumbing. In 2017 it was the initial coin offering — a story about democratized capital formation that turned out, for a season, to be a story about retail liquidity exported to a handful of exchanges. In 2020 it was DeFi Summer, when I was running DeFi Digest and we all learned to speak the language of liquidity pools and impermanent loss. That cycle had a genuine technical core — automated market makers really did rewrite market structure — but its emotional engine was the idea of financial sovereignty, and the two never fully reconciled. I wrote a piece then called "Impermanent Loss as Social Contract" that reached 200,000 readers, and the reason it resonated was not the math. It was that people wanted the plumbing to mean something.

In 2021 the NFT boom gave us the cleanest narrative of all: provenance as culture, code as canvas. I built an entire platform, ArtChain Chronicles, around the intersection of digital art and cryptographic ownership, and I interviewed twenty artists and ten protocol founders for a series I called "The Soul of the Token." The art was real. The culture was real. What was not yet real was the durable market, and when it evaporated, so did a great deal of the language we had used to describe it.

Then came the Layer 2 wars, and with them the most instructive myth of the modern era: that scaling is the same thing as growth. Between 2021 and 2024, dozens of rollups raised billions on the promise that cheaper blockspace would summon users. It summoned, instead, a diaspora. By early 2026 there are more than sixty production rollups in the Ethereum ecosystem alone, and I would be generous if I said the active-user base across all of them had doubled since 2023. What actually happened is that a single, scarce pool of liquidity and attention was sliced thinner and thinner, until each chain could claim a thriving community and none could claim a thriving market. This is not scaling. It is fragmentation wearing scaling's coat.

I learned to read failure modes the hard way. During the 2022 collapse of Terra, my inbox filled with readers who had lost real money, and my own confidence took a hit I did not expect. What pulled me out was curiosity — the compulsion to understand the structure beneath the story. I started a project I called the Post-Mortem Anthology, documenting the breakdowns of thirty major protocols and interviewing fifty veterans, and what I found across all of them was the same skeleton: a narrative that had outrun its mechanism, held aloft by leverage until the leverage ran out. Restaking, in 2023 and 2024, added a new chapter to that skeleton — security as a product you could rent, and then re-rent, and then re-rent again, until the word "security" had stopped describing a property and started describing a yield.

And now, in 2026, we have arrived at the AI-agent cycle, which carries the most seductive premise of all: that the next billion users will not be human, that machines will transact at a volume no retail cohort could ever match, and that whoever hosts those machines will inherit the next decade of fee revenue.

I want to be precise about where that premise breaks, because it breaks in a place most people are not looking.

Core: the anatomy of an indifferent user

To understand why agent adoption can be net-negative for a chain's token, you have to understand what an agent actually needs, and it is a shockingly short list.

An autonomous agent — whether it is a trading bot, an inference broker, a data-marketplace participant, or the kind of machine-to-machine service that the x402 standard was built to settle — needs three things from a blockchain. It needs finality it can verify without trusting a counterparty. It needs a cost structure it can predict at the margin, because it is making millions of decisions and cannot afford variance. And it needs a payment rail that its counterparty will accept. That is the entire specification. Notice what is not on the list: brand loyalty, community, governance participation, token alignment, or any of the soft variables that a chain's narrative depends on.

The x402 standard, which revives the long-dormant HTTP 402 status code as a native payment negotiation layer, is a useful lens here. When a machine requests a resource and receives a 402, it does not open a wallet app and weigh its options. It looks up the accepted payment method, signs a stablecoin transfer, and retries. The entire interaction is designed to be boring. And boring is exactly what kills a token narrative, because tokens are priced on the expectation of scarcity — of blockspace, of attention, of access — and agents are in the business of eliminating scarcity wherever they find it.

I have seen this pattern before, in a smaller venue. In 2023 I consulted on the audit of a solver network — one of the intent-based systems that fill user orders by competing in an auction. The engineers had built something genuinely elegant, and their go-to-market assumption was that solver activity would drive demand for their token. What the data showed eighteen months later was the opposite. The solvers routed around the token entirely. They settled in stablecoins, they paid for compute in fiat-denominated credits, and the token survived only as a governance artifact that nobody with real volume bothered to use. The activity was real. The value capture was fiction. The first law of the agent economy is that activity and value capture are no longer the same thing — and most chains have not yet noticed that the two have divorced.

The identity problem compounds the economics. ERC-8004, the emerging standard for trustless agent registration, exists precisely because an agent's counterparty needs to know that the machine on the other side of a transaction is who it claims to be, and that it will still be there tomorrow. This is a genuinely new requirement. Human users carry reputations, legal identities, and the diffuse accountability of a community. Agents carry none of these by default, which is why the identity layer is being built at all — and why it matters to the value-capture question. The entity that controls agent identity controls a chokepoint that agents cannot route around. In most architectures, that chokepoint is not the rollup. It is a registry, a reputation system, or a staking-based attestation: a thin layer sitting above the settlement chain, capturing value the chain itself never sees.

This is where the Layer 2 problem becomes acute. A rollup's business model, as originally conceived, was a spread: buy blockspace from Ethereum, sell it to users at a markup, and let the sequencer keep the difference. EIP-4844, the blob upgrade that went live in 2024, collapsed the cost of that input data by orders of magnitude — a genuine engineering triumph — but it also collapsed the markup that rollups had been living on. When the cost of your raw material falls toward zero and your customers are indifferent to your brand, your margin does not compress gently. It evaporates.

Now add agents. An agent does not choose a rollup the way a person chooses an app. It chooses the cheapest route that satisfies its constraints, and it re-chooses every time those constraints shift. This is not disloyalty; it is correct engineering. A rational agent will split its activity across five chains in a single day if the price signals tell it to, which means that a chain's "agent adoption" metric can spike to an all-time high in the same week that its token hits an all-time low — exactly the divergence I watched on March 4. The transaction count is real. The demand it represents is real. What is missing is any mechanism by which that demand is forced to pay for the token.

Mapping the chaotic beauty of market sentiment is normally my favorite part of this job, but sentiment here is a lagging indicator of a structural problem. Retail sees rising transaction counts and assumes rising value. The data sees rising transaction counts and falling fee capture, and understands that the two are not in tension — they are the same phenomenon. Cheap blockspace attracts high-volume, low-margin users. High-volume, low-margin users generate impressive charts and negligible revenue. The chart is the marketing. The revenue is the business. They have stopped pointing in the same direction.

To test the divergence I have been describing, I spent the first quarter of 2026 building a rough instrument. I pulled transaction-level data from a dozen rollups, filtered for wallets exhibiting agent-like behavior — high frequency, low value, no governance interaction, sponsored gas — and compared the growth in that activity against the growth in fee revenue captured by each chain's native token. Across the sample, agent-like transactions grew by a median of roughly ninefold over eighteen months. Fee revenue attributable to those same transactions, net of paymaster subsidies and sequencer rebates, grew by less than 40%. In three of the twelve chains, it declined. That is the shape of a growth story that is quietly a margin story in reverse.

Latency deserves a closer look, because it is the dimension agents optimize that humans barely notice. A person tolerates a two-second confirmation; a machine arbitraging a price discrepancy across venues is racing other machines in milliseconds. This pushes agent activity toward centralized sequencers with soft pre-confirmations, and away from architectures that prioritize decentralization at the cost of speed. It also reopens the MEV question in a new form: when the fastest participants are machines with no legal identity, the extractable value they capture is no longer a market inefficiency to be taxed — it is a structural transfer from the chain's users to a fleet of indifferent optimizers. The chain that hosts the agent captures the congestion; the agent captures the value.

There is a second-order effect that almost nobody is pricing, and it concerns who actually wins when agents commoditize blockspace. If the marginal agent is indifferent between sixty rollups, then the rollup layer converges toward a commodity — differentiated by cost and latency, not by narrative — and the surplus migrates to two places. The first is the settlement layer underneath, because every rollup still posts its data and its proofs to Ethereum, and Ethereum's security budget is the one input no rollup can route around. The second is the stablecoin issuers, because the agent economy is being built, almost entirely, on dollar-denominated rails. When an agent pays for inference or settles a micro-debt, it is not doing so in a volatile governance token. It is doing so in a stablecoin, because a machine optimizing an objective function cannot afford the variance.

The restaking economy offered a preview of how this surplus migrates. Restaking was sold as a way for a chain to rent security rather than build it, and for a season the token mechanics looked compelling. What the data eventually revealed was that the value accrued to the restaking middleware and the underlying settlement layer, while the chains that "bought" security paid a rent that showed up as a permanent line item on their economics. Agents will do to blockspace what restaking did to security: turn a scarce resource into a metered utility, and route the surplus to whoever owns the meter.

The strategic error underneath all of this is a category error about customer lifetime value. A human user, at least in the model most chains still carry, arrives with an identity, develops habits, accumulates a reputation, and — crucially — can be monetized through a token that appreciates as the community grows. An agent arrives with none of that. Its lifetime value to the chain is the sum of the fees it pays before it routes elsewhere, which is to say, very little, for a very short time. Chains that measure success by user count are importing a metric from the human era and applying it to a user that has redefined the word.

This is the quiet part of the story, and it is worth sitting with. We spent a decade arguing about which chain would win, and the answer emerging from the agent economy is that the chain may not be the unit of competition at all. The competition is between settlement assurances and payment rails, and the chains are increasingly just the venue where those rails happen to run. Following the thread from code to culture leads somewhere uncomfortable for anyone whose thesis is a ticker: the code is converging, and the culture is being outsourced to machines that do not have one.

What is actually being repriced, beneath all the tickers, is the narrative premium — the extra value a token carries because it represents a story about the future. Agents do not pay for stories. They pay for blocks, at a price they can compute. Every quarter that agents make up a larger share of activity, the narrative premium compresses, because the marginal buyer of the underlying service is no longer a person who believes in anything. That is not a bearish call on the technology. The technology is working. It is a bearish call on the practice of pricing infrastructure like it is culture.

I should be careful not to overstate the speed of this. Agents are not yet the majority of on-chain activity, and human users still anchor the demand for most applications. But the growth rate matters more than the level, and the growth rate of agent-driven transactions has been running at multiples of the human rate for six consecutive quarters. When a new class of user grows faster than the incumbent class, it eventually sets the price. And the price this class is setting is low — deliberately, structurally, permanently low.

The RWA conversation is a useful contrast, because it reveals how differently institutions treat the same infrastructure. For three years, the loudest voices in this industry have promised that tokenized real-world assets would bring trillions of dollars of institutional capital on-chain. What has actually arrived is a much narrower thing: a handful of tokenized treasury products, most of them distributed through permissioned environments that look nothing like a public mempool, and almost none of them settling on the rollups that spent their treasuries courting them. Traditional institutions do not need your public chain. They need settlement finality, legal clarity, and a counterparty they can sue. The public-chain part is a convenience, not a requirement, and the moment it becomes inconvenient they will route around it — exactly the way an agent routes around a rollup whose fees have ticked up. The institution and the agent share a temperament, and it is not a sentimental one.

The same logic is quietly dismantling the "Bitcoin Layer 2" category. Most of the projects wearing that label are Ethereum infrastructure in a different costume, built on bridges and multisig custody and a settlement guarantee that the Bitcoin base layer never actually provided. The real Bitcoin community does not acknowledge them, and it does not need to. They exist to capture a narrative, not to serve a settlement function — which is exactly the pattern agents will punish, because an agent cannot be persuaded by a narrative it cannot verify in a block header.

Contrarian: the deflationary arrival

The consensus narrative of 2026 is that AI agents are the demand event that finally justifies the infrastructure build-out. I want to argue the opposite, or at least a sharper version of it: that the arrival of agents is, for most Layer 2 tokens, a deflationary event dressed as a growth story.

Consider what has to be true for agent adoption to be bullish for a rollup's token. Agents have to be sticky — they have to prefer your chain for reasons that persist. Agents have to pay fees that accrue to the token, not to a paymaster, not to a solver, not to a sequencer subsidy that the foundation funds out of its own treasury. And agents have to be willing to hold the token, or at least to create demand for it through some mechanism other than speculation. In practice, none of the three holds. Agents are the opposite of sticky. Their fees are abstracted by design, because the entire point of account abstraction and paymasters is to make gas invisible. And they have no reason to hold a governance token for a chain they treat as interchangeable.

What agents do have a reason to do is drive down the price of the thing the token represents. Every agent that routes to the cheapest chain applies downward pressure on fees. Every paymaster that sponsors gas removes a revenue line. Every solver that internalizes an order removes a fee that would otherwise have reached the sequencer. The system is working exactly as designed, and the design is deflationary for everyone in the middle. The surplus does not vanish; it migrates — down to the settlement layer, and sideways to the stablecoin rails.

There is a tempting counter-argument: that volume begets liquidity, liquidity begets applications, and applications beget demand. This is the standard reflexive loop, and it held in DeFi Summer because the users were humans with speculative appetites who could be captured by a token. Agents cannot be captured by a token, because they do not have appetites. They have parameters. You cannot build a community around a user that does not experience FOMO.

The honest version of the bull case is narrower than the one being sold. It is that a small number of chains will become the default settlement venues for agent activity — not because agents love them, but because they achieved a cost and reliability advantage that was expensive to replicate — and that everything in between becomes a commodity. That is a real and investable thesis. It is just not the thesis that sixty rollups are currently marketing, and it is not the thesis that the tokens are priced for. The blind spot is not that agents are coming. It is that everyone is modeling them as users when they are actually competitors — competitors for the fee, the spread, and the margin that the current business model depends on.

Takeaway

If the agent economy is the next cycle, then the question that matters is not which chain has the most agent transactions. It is which layer of the stack can charge for something an indifferent machine cannot route around. Finality is one such thing. Trust-minimized settlement is another. Dollar liquidity is a third, and it may turn out to be the largest of them all. Everything above that — the branded rollup, the points program, the governance token with no fee flow — is a layer a machine will happily skip.

I keep coming back to that wallet on March 4, and to the 4,118 transactions that moved almost nothing, paid almost nothing, and meant almost nothing to the chain that hosted them. Artifacts of a new digital renaissance tend to look like this at first: small, strange, and easy to mistake for noise. But the renaissance this time is not being painted by humans. It is being compiled by agents, and agents do not sign their work.

The next narrative will not announce itself with a price chart. It will announce itself in the cold arithmetic of who still gets paid when the users stop being people. Watch the fee lines, not the transaction lines. One of them is the story. The other is the ghost.