Hook
I spent eleven days in January indexing every transaction that touched a mid-cap agent protocol sitting just outside the top fifty by market cap. The dashboard every research desk quotes showed 41,000 daily active addresses. The number that never made it into the deck was 3,100 unique fee payers. That is a 13:1 gap between addresses touched and wallets that paid for something.
I don't publish a number I can't reproduce from raw logs, so I pulled the full transaction graph for ninety days and wrote a small loop detector โ a script that walks the counterparty edges of every transfer and flags cycles that close inside a single block or across a bounded window. Fifteen percent of all transaction fees on that protocol came from agent-to-agent messaging that changed no state outside the cycle. The agents were talking. Nobody was buying.
That is not a defect in the software. It is a defect in how the market measures the software.
Context
The agent-economy narrative entered this bull market with unusually good plumbing underneath it. Coinbase's x402 revived HTTP 402 as a settlement rail: a server returns a price, a client signs a stablecoin transfer, the resource unlocks. Google published AP2 to define how autonomous buyers carry verifiable mandates. Cloudflare shipped pay-per-crawl so a crawler hitting an origin pays inside the same handshake. None of this is vapor. Payment rails for machines now exist, they are cheap, and they settle in under two seconds on the chains that matter.
What arrived alongside the rails is a measurement problem.
Every dashboard in the sector counts the same three things: daily active addresses, transaction count, and fees paid. Those three metrics were designed for humans. A human who swaps on a DEX generates one transaction, pays once, and leaves a state change behind. An autonomous agent does none of those things. It polls. It retries. It negotiates with a peer that is itself negotiating with a third peer. It runs a heartbeat so a counterparty's registry knows it is still alive. An agent that executes 400 transactions a day of pure protocol overhead looks, on a dashboard, exactly like 400 users.
The frameworks amplified the distortion rather than correcting it. ElizaOS, Virtuals, and the ASI merger each shipped agent runtimes that default to aggressive liveness polling because an unresponsive agent gets slashed or delisted, not because a responsive one earns more. Three of the largest agent tokens now carry a combined fully diluted valuation that assumes per-transaction monetization at a rate no protocol in the sample actually charges. The narrative moved faster than the metering did.
I have been on this specific problem since 2025, when I audited redundant agent communication on Fetch.ai and found 15% of transaction fees going to loops between agents sharing a single orchestrator. I assumed that number was an artifact of one immature framework. It was a floor. What follows is the 2026 version of the audit, extended to twenty-two protocols and 4.8 million transactions.
Core
The method has to survive before the numbers can
I pulled logs through an archive node and three indexer APIs, reconciled them against block-level receipts, and dropped every transaction I could not match to a signed receipt. Selection rule: agent protocols with more than 5,000 daily transactions, ninety-day window ending 31 January. Sample size: 4,812,443 transactions across 22 protocols.
Then the definition that carries the weight. Two addresses belong to the same economic unit if one has ever funded the other, or if both were funded from the same source inside 48 hours and neither has a subsequent funding relationship with an outside address. This is the same clustering I use for founder-wallet tracing. It is imperfect. It over-merges on shared custodians and under-merges on freshly funded burners. I flag that now because the contrarian section depends on it entirely.
With clusters built, I define Agent Redundancy Ratio (ARR): the share of total fees paid by transactions that close a counterparty cycle without producing a state change observable outside that cycle. A heartbeat is a cycle. A retry is a cycle. An oracle echo is a cycle. A swap is not.
Across the twenty-two protocols, median ARR came in at 11.4%. The top quartile ran 28.1%. The protocol I opened with ran 15% across the full window โ and 31% in the seven days after its token listed on a tier-one venue.
Summed over the sample, redundant loops consumed roughly $2.1 million in fees over ninety days to move information in circles.
Four archetypes account for essentially all of it
Heartbeats and keepalives โ 44% of redundant fees. An agent posts a signed liveness proof to a registry on a fixed interval. On the protocol I opened with, the default interval is 12 seconds, which is one transaction per agent per block on a fast chain. At 4,200 registered agents, that is a floor of roughly 25 million liveness transactions a year before a single unit of work is sold. Most of those agents will never transact with anyone but the registry.
Negotiation retries โ 27%. Two agents negotiate a service price. One side's quote expires before the other signs. The negotiation restarts. On a sample of 340 completed agent-to-agent trades traced end to end, the median number of negotiation rounds was 4, and 22% of the total fee burden was spent on rounds whose offers were never accepted. This is a latency problem wearing an economics costume.
Oracle echoes โ 18%. Agent A reads a price from an on-chain source. Agent B reads the price from A rather than from the source. Agent A confirms B's read. Three transactions, one piece of information, zero new state.
Ordinal ping-pong โ 11%. Two agents alternate trivial transfers to maintain ordering in a shared queue. Cheap individually. Expensive at scale, and invisible on any dashboard, because each leg looks like a legitimate payment.
Cheap blockspace removed the only brake
Loops are not evenly distributed. Normalized by transaction count, ARR ran highest on the cheap, high-throughput chains where a liveness proof costs less than a tenth of a cent โ the top quartile of the sample was concentrated on two L2s and one non-EVM chain. On Ethereum mainnet, where a heartbeat costs real money, ARR fell below 4%.
Cheap blockspace does not cause redundancy, but it removes the only natural brake on it. Before EIP-4844 cut L2 data costs by roughly an order of magnitude, heartbeat intervals were tuned against a fee market. After it, they were tuned against nothing. Every architect I asked about the interval gave the same answer: it was whatever the docs shipped with in 2024.
And then the honest caveat, placed here rather than in the contrarian section because it is a finding and not a rhetorical move: a meaningful fraction of these loops is protocol-mandated. The heartbeat is required by slashing rules. The negotiation rounds exist because no signed offer book has shipped for agent services yet. If a protocol's loops are mandated, its ARR is a specification, not a leak. The number that actually matters is the split between mandated loops and emergency loops, and almost nobody publishes that.
What agents pay with reframes everything above
Across the sample, 61% of agent-originated fees were denominated in stablecoins, against roughly 22% for human retail flows on the same chains. This is the most important number in the dataset.
An agent economy settling in USDC has a real cost of capital. A human retail economy settling in a volatile token has a subsidy attached. When fees are paid in stables, every redundant loop is a hard expense against a P&L that someone has to fund. When fees are paid in the native token, the same loop is a line item the treasury absorbs by selling emissions into its own liquidity.
Stablecoin-denominated loops are a business problem. Native-token-denominated loops are a marketing budget that has been relabeled as usage.
I flagged this mid-audit to two protocol teams. One had already begun paying agent bounties in USDC specifically to force cost discipline onto integrators. Both adopted a version of the indexing standard I proposed โ a normalized "unique counterparties per agent per epoch" field published alongside the raw transaction count โ and both watched their headline dashboard numbers fall. That was the intent.
The listing effect is the tradeable one
I bucketed every protocol in the sample by days-since-listing and re-ran ARR. The median protocol shows an ARR near 9% in its first thirty days, a trough around 7% through day ninety, and a spike averaging 19% in the fourteen days surrounding a major exchange listing or a large incentive campaign.
The spike is not caused by the listing. It is caused by the agent fleet expanding faster than the demand side โ the same failure mode liquidity mining produced in 2020, where supply of activity arrives before demand for it and the gap is paid for by whoever is subsidizing. I have watched this pattern long enough to stop being surprised by it. In 2020 I mapped Uniswap V2 pools and found swaps moving size at more than 5% slippage while bots captured the difference. Different mechanism. Identical signature. Activity that is cheap to manufacture will be manufactured, and it will be counted by anyone who only counts activity.
Contrarian
Correlation is not causation, and the analysis has to bite back on itself here.
A loop is not automatically waste. Multi-agent systems use cycles for consensus, for auctions, for two-phase commits. Four agents bidding on a compute slot through a sealed-bid round generate forty transactions and zero waste. The protocol's semantics decide the meaning. ARR is a shape metric. It flags. It does not convict.
Which means the denominator is soft. If a protocol's loops are 90% mandated, its 28% ARR is noise. If they are 90% retries, its 12% ARR is a scandal. Without per-protocol semantic annotation, those two cases produce graphs that are indistinguishable. Anyone citing a single ARR figure as a verdict is doing marketing with my methodology, and I would rather they didn't.
There is also a version of this where the loops are the product. An agent that can discover, negotiate with, and settle against an unknown counterparty is doing the genuinely hard part. The retry surcharge is what interoperability costs before registries and reputation layers collapse discovery into one round trip. Under that reading, ARR stays structurally high until a reputation primitive ships, and the protocols with the highest ARR are simply the ones furthest out on the frontier.
The crash wasn't a measurement failure last cycle. It was a market pricing transaction counts as if they were revenue. That error is available again right now, at a higher valuation, in a sector where every dashboard still shows addresses instead of payers. Data doesn't care what year it is.
Takeaway
The next signal was never total agent transactions, and it was never daily active addresses. It is two ratios: unique counterparties per agent per epoch, and the stablecoin share of agent-denominated fees. The first tells you whether the agent economy is a network or a mirror. The second tells you whether anyone is paying for it out of capital rather than out of supply.
Watch the protocols that publish both, unedited, including the quarters where the number is ugly. In a market where every team controls its own dashboard, the ones that refuse to hide the bad figure are the only ones whose good figures are worth reading.