Hook
Last Tuesday I spent ninety minutes pulling the role graph of a rollup that closed an $85 million raise in March. What I found wasn't a hack. It was worse. It was the design.
DEFAULT_ADMIN_ROLE on the rollup's core contracts belongs to a proxy admin contract. That proxy admin is owned by a 4-of-7 multisig. Three of the seven signers are foundation employees. Two are the same two funds that led both the seed and the Series A round. Of the remaining two, one is a paid advisor to a portfolio company held by one of the leads. The timelock is 48 hours. Sitting above the timelock is a guardian role, held by a single externally owned account, with unilateral power to cancel any queued proposal before execution.
The token carries a $2.1 billion fully diluted valuation. The website says "community-governed."
I am not naming the project, because the project is not the story. Run the same query against the last twenty rollups that shipped a token and you will find the same skeleton wearing different branding. Decentralization in this cycle is a marketing department, not an architecture.
Context
Be precise about what changed, because this bull market is doing what bull markets do — converting expensive engineering problems into cheap narratives.
More than sixty production layer-2 networks now settle to Ethereum. Another thirty are funded and announced. The pitch is structurally identical: cheaper execution, faster finality, and a roadmap that terminates in "progressive decentralization." That last clause has done an enormous amount of work. It is the crypto equivalent of "we'll fix it in post."
In 2020 I spent two weeks reverse-engineering the Uniswap v2 bonding curve and published a series arguing that centralized exchanges had become structurally obsolete because of MEV extraction. The piece was wrong in one direction and right in another. CEXs did not die. But the AMM became the load-bearing wall of the entire market. The lesson was not about AMMs. It was that the market prices the mechanism it can see and ignores the mechanism it cannot. In 2020 nobody priced MEV. In 2026 nobody prices the admin key.
In 2022, when UST depegged and the timeline filled with guesswork, I spent four hours tracing the Luna Foundation Guard's reserve diversification instead of writing about the price. The failure was mechanical — a stability mechanism funding its own defense with an asset whose value was derived from the mechanism itself. Institutions cited that breakdown for months. It changed how I run a desk: verify first, publish second. Our churn rate dropped by half, and subscription retention went up.
Go back further and the pattern holds. In 2017, as a junior analyst doing fast audits of ICO whitepapers, I found a reentrancy bug in the Zcoin contract hours before its token generation event. Roughly $2 million in retail deposits did not get destroyed that week. That set the standard I still run on, and it is why I read contracts before I read roadmaps.
So here we are again. Except this time the flaw is not a reentrancy. It is governance.
Core
Start with the contracts, because contracts do not lie and landing pages do.
Every one of these networks ships an upgradeable contract suite. That is not a criticism — upgradeability is rational for a young protocol. But it has a price, and the price is that "code is law" becomes "code is law until the admin calls upgradeToAndCall." The proxies are UUPS or transparent. The storage layouts are versioned. The audits are real and the auditors are competent.
And every one of those audit reports contains a section titled something like "Centralization Risks," and inside that section, there is a line that reads: Acknowledged.
I have read dozens of these reports this year. The taxonomy is remarkably stable. AccessControl grants DEFAULT_ADMIN_ROLE to a deployer-controlled address. A PauserRole can halt deposits and withdrawals. A ProposerRole exists but is gated by a token vote — which sounds reassuring until you notice that the vote only proposes, and the multisig executes. And the timelock, that sacred 48-hour window meant to give the community time to exit, is cancellable.
That cancellation primitive is the whole ballgame. A timelock that can be cancelled is not a timelock. It is a notification.
Walk the lifecycle of a real governance proposal on one of these networks and the theater becomes legible. There is a forum post, then a temperature check, then a Snapshot vote — off-chain, non-binding, free to cast, and in many cases legally decorative. Then, if the temperature is high enough, an on-chain proposal. Then the timelock. Then execution. At no point in that chain is the foundation obligated to honor the outcome, because the execution path runs through a multisig that is not bound by the Snapshot result at all.
Quorum tells you the rest. On the networks I have sampled, participation on binding proposals sits between two and five percent of circulating supply, and a meaningful share of that is delegated to fewer than twenty addresses. Vote-buying is trivial in a market where the governance token is liquid and lendable — you do not even need to buy the token, you can borrow it, vote, and return it before the next block. Delegation markets have made this frictionless. Code is law, but audits are mercy — and nobody is auditing the voter, only the vote.
I keep returning to something an editor told me in 2017, when I published the Zcoin warning: the difference between a vulnerability and a feature is whether the documentation mentions it. The docs here do not mention the guardian. The docs mention a "security council with emergency powers." Security council is the polite term. It is also, technically, accurate. It is just not the word that reaches the headline, and it is definitely not the word that reaches the valuation model.
Here is the uncomfortable arithmetic. When a token's price is derived from a governance narrative but its upgrade rights are derived from a multisig, you are not holding a share of a protocol. You are holding a call option on the goodwill of seven people. The only governance right retail actually exercises is the exit, and the exit is priced in real time.
Now zoom out, because the governance problem is downstream of a bigger one.
In 2020, "liquidity fragmentation" meant a handful of venues. Today it means dozens of execution environments with separate state, separate bridges, separate sequencers, separate gas tokens, and separate incentive programs, all bidding for the same depositor base. Each new bridge is also a new trust surface — we have watched this movie before, and the honest summary is that every additional hop multiplies attack surface without multiplying users.
The measurement is where the story gets interesting. When you read "Total Value Locked," you are almost never reading net new capital. You are reading the same dollar counted at multiple points along a route. A user deposits ETH on Ethereum, bridges to a layer 2, deposits the receipt token into a lending market, then posts that LP position as collateral somewhere else — and a single ETH now appears in four TVL figures. The pool remembers what the ticker forgets.
So when a rollup announces it crossed $1 billion in TVL, ask the only question that matters: net or gross? Nearly always gross. Nearly always bridged in, not native. Nearly always incentivized, not organic. And increasingly, nearly always about to leave.
The scarce resource in this market is not blockspace. I have been writing that since the DeFi summer, and it gets truer every cycle: liquidity does not die, it migrates — and it migrates toward whoever signs the checks. Sequencer revenue funds the incentive programs, and on most networks that revenue is a fraction of the incentives themselves. Which means the liquidity is rented, not owned, and the lease is denominated in a token that the same four keys can mint. That is not a criticism of the token model. It is a description of a reflexive loop that has never been stress-tested in a sustained drawdown.
Restaking added another layer without resolving the first one. Capital that was already counted once is now counted again as security, and the correlation risk between the underlying asset and everything secured by it compounds quietly. We keep inventing new ways to reuse the same collateral, and then describing the result as growth.
Which brings us to the part of this bull market that is barely being audited at all.
The agent thesis. My own framework, published last year, puts autonomous on-chain agents at roughly 60% of transaction volume by 2027. I still believe that number. But the interesting variable is not the volume. It is where the volume settles.
Agent payments need different primitives than human payments. Session keys. Paymasters that sponsor gas. Deterministic execution. Revert semantics that do not cascade. EIP-7702 lets an agent operate through a delegated EOA, which is elegant and also means the security model now includes every key-management decision made by the developer who wired it up. An agent routing a thousand micro-transactions an hour does not care about validator count. It cares about p99 latency, fee predictability, and whether a reorg can invalidate the position it built three seconds ago.
So put yourself where the agent is. You have a choice of rails. One is credibly neutral, has a wide validator set, and finality measured in minutes. Another has a single sequencer, sub-second finality, and a fee market so predictable you can model it to the basis point.
The agent picks the second one. Every time. Not because it is decentralized. Because it is reliable.
And here is what should make governance tokens uncomfortable: from the agent's perspective, the four keys are not a liability. They are an operational guarantee. If something breaks, someone can fix it. Agents do not want immutability. Immutability is a bug that reverts.
The truth is hidden in the gas fees. Pull the fee distribution on the top rollups this quarter and the pattern mirrors what I found in CryptoPunks wallet data in 2021 — a small cluster of counterparties driving a disproportionate share of activity. In 2021 it was whales positioning ahead of floor prices. In 2026 it is bots paying priority fees for certainty. In both cases the headline metric was a lagging indicator dressed up as a leading one.
Before the bots, there were sybil farms. The points programs, the retroactive airdrop hunting, the wallets that exist only to farm and exit — those were the rehearsal for agent volume, and the industry treated them as a nuisance rather than a preview. When the same behavior arrives with better engineering and no sleep schedule, we will call it adoption.
The Bitcoin parallel is instructive here, and uncomfortable. I have argued consistently that Ordinals injected something Bitcoin badly needed: fee revenue. Without the inscription wave, the security budget conversation gets very sharp, very fast, because subsidies halve on a schedule and fees have to fill the gap. The inscription era proved that demand for block space can materialize from directions the core community did not anticipate — and that narrative, not just utility, is a load-bearing component of a security model.
Sit with that, because it cuts both ways. If a chain's security depends on narrative-driven fee demand, then narrative is infrastructure. And infrastructure can be deprecated. When the mint wave cooled, fees compressed and blocks got lighter. Nobody panicked. But the rehearsal happened, and the lesson is that we audit Solidity for reentrancy and we do not audit incentives for reflexivity.
Entropy increases until someone audits it. Nobody has pointed that lens at the flywheel.
Contrarian
Here is the angle the market is missing, stated as plainly as I can manage: the agent economy will push layer 2s toward more centralization, not less — and the industry will let it happen, because it will be efficient.
Every decentralization roadmap I have read this year has the same shape. Sequencer, then shared sequencer, then a decentralized sequencer set, then based sequencing. Each stage is gated on a research milestone that has not shipped. Meanwhile the agent volume arrives now. Agent volume does not tolerate probabilistic finality. So teams will keep the fast path and ship the decentralized path "soon," and the vote to extend that timeline will pass with three percent turnout.
The tell is already in the marketing. Based sequencing is being sold as a return to Ethereum's security. What it actually does is outsource ordering to the L1 builder market — which is itself concentrated in a handful of builders and relays. The decentralization did not multiply. It moved.
There is a second-order effect that anyone holding a governance token should sit with. If agents become the majority of volume, they become the majority of fee payers. If they are the majority of fee payers, protocols optimize for them. Optimizing for agents means predictable gas, and predictable gas looks less like a priority-fee auction and more like a subscription. The human retail user becomes the marginal, cross-subsidizing participant in a system whose governance token was minted to represent them.
Speculation is just data with a heartbeat, and the heartbeat of this cycle is a bot arbitraging five rollups simultaneously. Retail is the exit liquidity for an algorithm that never sleeps and never reads the forum. That is not cynicism. That is the order book.
Volatility is the tax on uncertainty. This cycle, most of the uncertainty is not in the price. It is in the upgrade path.
Takeaway
Watch different things. Not the TVL headline. Not the validator count. Watch the role table. Watch whether the guardian veto is removed on schedule, or whether the schedule gets quietly extended by a proposal that six addresses bothered to vote on.
Rewriting the rules before the bug writes them is the only version of decentralization that survives contact with agent-driven order flow. The chains that do it will look boring for a quarter and dominant for a decade. The rest will run a $2 billion narrative on four keys and a marketing budget.
The question I would put to every one of those teams is a single sentence long: when the agent volume arrives, who signs the upgrade?
