In the second quarter of 2025 I rebuilt the agent simulation I first stood up in January — the same model I used to convince our investment committee to underwrite data-availability and identity infrastructure before the rest of the desk understood why. The output was not the number I expected. Machine-to-machine transaction volume did climb, and it climbed faster than my own base case: agent-originated calls ran roughly 200% above the prior year's baseline. But the number that mattered was not the volume. It was the failure curve. At a 92% single-step success rate — generous, by current frontier-model standards — a ten-step task completes end-to-end about 43% of the time. At fifteen steps, you are under 29%. The agents Big Tech is pitching as the next digital interface do not fail because they are unintelligent. They fail because reliability compounds against them.
Ignore the keynote language. Look at the failure curve. When Microsoft, Google, Anthropic, and OpenAI converge on the same marketing frame inside the same two quarters — "the agent is the new interface" — the interesting question is never whether the technology works in a demo. The interesting question is what the frame is designed to hide. What it hides is a distribution war dressed as a product launch.
That is the whole article. Everything below is the proof.
The proximate signal is thin by design. A trade brief lands: Big Tech is pitching AI agents as the next major digital interface, with three words doing the work — trust, interoperability, efficiency. No named companies. No numbers. No quotes. A press release wearing the clothes of analysis.
Thin signals are not empty signals. The verb is the tell. "Pitches" is a marketing verb, not an engineering verb. When a sector moves from building to pitching, it has crossed a boundary — from capability demonstration into market education and land acquisition. The DeFi summer of 2020 taught me to read that transition precisely. When a protocol stops publishing audits and starts publishing APY banners, the yield curve has already inverted. Volume without conviction is just noise, and the loudest pitches arrive after the smartest capital has already positioned.
Set the macro map first, because the interface debate does not float free of liquidity. We are in a consolidation regime — a sideways tape, no clear direction, rate expectations oscillating with every data print. In this environment capital does not chase growth stories; it chases optionality. AI agents are the cleanest optionality story on the board, because they promise to expand the total addressable surface of software without requiring new hardware, new users, or new demand. They monetize the install base. That is why the pitch is happening now, and not two years ago.
The three keywords are not filler. They are a competitive map. Trust is the security-and-permission layer — who the agent is, what it may do, whether it left a trail. Interoperability is the protocol layer — how agents talk to tools and to one another. Efficiency is the ROI layer — the language you use when you are selling to a CFO rather than a CTO. Read together, they describe not a product but an ecosystem, and ecosystems are won by whoever controls the entry point.
Which brings us to crypto, and to why a crypto-native desk should care about a story that contains no crypto. The agent economy is a machine economy, and a machine economy needs three things centralized stacks are structurally bad at providing: permissionless settlement, portable identity, and verifiable execution. Crypto already ships all three, badly. The question is not whether agents will touch blockchains. My simulation says they already will, at scale. The question is whether crypto captures the interface or is demoted to the rail beneath it.
One more read of the source matters. The brief comes from a crypto publication, yet contains no crypto. That absence is itself a signal — a crypto desk flagging a pure-Big-Tech story usually means the intersection trade is forming but not yet public. When the crypto press starts covering the agent pitch without mentioning blockchains, the blockchains are being priced quietly.
Start with the reliability cliff, because it is the most mispriced variable in the entire narrative. The industry pitches agents as an interface upgrade — a nicer way to get things done. That framing is wrong. An interface that answers is a tool; an interface that acts is an agent, and the moment you move from generating text to executing operations, you inherit every failure mode of a distributed system: error accumulation, tool-call hallucination, state-management drift, cross-session memory inconsistency, cost blowout. None of these are model problems. All of them are systems problems, and systems problems do not dissolve when the next checkpoint drops.
The binding constraint on the agent economy is not intelligence; it is the arithmetic of compounding failure. A 92% single-step success rate sounds like near-perfection. Run it through ten dependent steps and you are at 43%. Fifteen steps and you are at 29%. Real enterprise workflows are not five-step parlor tricks; they are twenty-, thirty-, forty-step chains with branch points and external dependencies. At forty steps, even a 95% per-step success rate yields an end-to-end completion probability of roughly 13%. This is why the industry's "POC graveyard" is not a rumor — it is a mathematical certainty. Pilots succeed in the lab because labs run short tasks. Production fails because production runs long ones.
I watched the same pattern in 2017, when I audited the reserve liquidity of five ICO treasuries with a Python script tracing Ethereum mainnet flows. Three of the five held under 5% of their claimed reserves in cold storage. The whitepapers were flawless. The tokenomics were elegant. The capital was not there. The lesson transferred directly: the narrative layer and the settlement layer are different systems, and only one of them survives a stress test. Illusions dissolve under stress testing. The agent narrative has never been stress-tested at forty steps, and the people pitching it know that.
Now the protocol layer, where the real war is being fought. Anthropic shipped the Model Context Protocol in late 2024 as an open standard for connecting models to tools. Google answered in April 2025 with Agent-to-Agent, a protocol for agents to coordinate with each other, backed by dozens of launch partners. OpenAI runs its own stack. Microsoft threads MCP compatibility through Copilot to inherit the Office and Windows distribution base. On paper this is a technical disagreement about message formats and capability discovery. In practice it is the same contest I have watched play out across Layer 2 for four years. The real difference between competing agent protocols is not architecture — it is which coalition convinces more builders to ship first. OP Stack versus ZK Stack was never a cryptography debate. It was a distribution debate wearing a cryptography costume, and the winner was decided by developer onboarding, not by proof systems. MCP and A2A will resolve the same way. The better protocol loses to the better coalition.
Here is the part the pitch hides. The interface is not where value is created; it is where value is captured. When the browser became the interface, the browser vendor captured the default search economics. When the app store became the interface, the store operator captured 30% of every transaction. If the agent becomes the interface, the agent operator captures the point of intent — the moment a human stops typing commands and starts stating goals. Whoever owns that moment owns the next decade of software distribution, and that is worth more than any single model's benchmark score. This is why "interoperability" is simultaneously the industry's stated goal and its sharpest weapon. A universal protocol is good for everyone in the abstract and fatal to whoever was counting on lock-in. Google and Anthropic push open protocols because open protocols convert research leadership into ecosystem gravity. OpenAI's tilt toward a closed loop is the same instinct inverted — protect the entry point, monetize the intent.
Follow the token, because the token is where the economics actually live. An agent task burns ten to one hundred times the tokens of a simple chat turn — multi-step reasoning, tool calls, result integration, retries. If agents scale, inference demand does not grow linearly; it grows non-linearly, and it concentrates on the inference side of the stack rather than the training side. That is a structural shift with two consequences. First, it makes latency the new scarce resource: a twenty-second agent response is unusable as an interface, so the winners will be whoever runs cheap, fast, low-latency inference, not whoever runs the largest model. Second, it exposes the unit-economics question no pitch deck answers: if a single agent task costs more to run than the human action it replaces, the efficiency story collapses. Efficiency is not a feature of agents; it is a threshold they must clear to exist. Follow the vector, not the hype — and the vector here points at inference cost per completed task, not at parameter counts.
This is where the crypto-native reading diverges from the mainstream one. The agent economy is a machine economy, and machines do not want user accounts, KYC forms, or monthly subscriptions. They want to pay per call, prove identity without a human, and settle instantly. Crypto is the only stack that natively provides all three, and the builders have already started shipping the primitives: machine-payable HTTP rails, on-chain agent identity registries, verifiable inference attestations, data-availability layers for agent state. In my 2025 simulation I modeled agent-driven load against gas markets and oracle feeds and found a predictable pathology — autonomous agents arbitrage latency and fee differentials faster and more ruthlessly than any human desk, which means oracle staleness and gas auctions become attack surfaces rather than infrastructure. The first crypto-native agent crisis will not be a hack; it will be a coordination failure in the fee market. That is not a reason to avoid the sector. It is a reason to own the parts of it agents cannot route around: settlement, identity, and data availability.
The trust layer deserves its own accounting, because "trust" is the first of the three keywords and the least examined. When an AI moves from generating content to executing operations, the risk surface migrates from saying the wrong thing to doing the wrong thing, and the harm is not linear — it is exponential. An agent that reads a web page or an email can be injected with malicious instructions. An agent with system or account permissions can act on them. An agent operating across applications widens the data-exposure surface with every integration it touches. The industry's own framing of trust reduces to three components: identity, authorization, and auditability — who acted, what they were permitted to do, and whether a durable record exists. Note what that triad actually is. It is a ledger problem. It is a verifiability problem. It is, whether the people pitching it admit it or not, the exact problem blockchains were built to solve, and the exact problem a centralized log file solves only until the answer carries financial or legal weight.
There is a second-order consequence the coverage keeps missing: the agent economy is a data economy, and data availability is the real chokepoint. Every agent action produces state — context, tool outputs, intermediate results, provenance — and that state must be cheap to write, cheap to read, and expensive to falsify. Centralized clouds do two of the three. They write cheaply and read cheaply. They falsify cheaply too, because the same operator that stores the record can edit it. For an interface that executes, that is not a feature; it is a liability. The agent era will be won by whoever makes agent state verifiable at cloud-scale cost, and that is a data-availability problem before it is an intelligence problem.
And here is the tell the crypto establishment is not ready for. The moment agents start interacting with DeFi, they will do to lending protocols what algorithmic traders did to equity markets in the 1990s — arbitrage every inefficiency until it disappears. Aave and Compound do not price interest by supply and demand in any meaningful sense. They price it by a governance-set curve — a utilization-based formula that is, at bottom, an arbitrary policy choice dressed as a market mechanism. Humans are too slow, too fragmented, and too polite to punish that arbitrariness. Agents are none of those things. An agent that can borrow across ten venues in a single block will treat every hand-tuned interest-rate curve as a mispricing, because that is exactly what it is. The protocols that survive the agent era will be the ones whose rate models are robust to machine-speed arbitrage, not the ones with the prettiest dashboards. Most are not ready. They were built for a world where the borrower was a human who slept.
The same logic applies, more uncomfortably, to Bitcoin. The ETF wrapper completed a transformation that began long before agents entered the picture: BTC is now a Wall Street instrument, held in custody by the same institutions that once derided it, traded by the same desks that price every other risk asset. Satoshi's peer-to-peer electronic cash is dead, and no agent will resurrect it. Agents do not want a bearer asset with a fifteen-minute settlement window and a volatile unit of account; they want programmable, instant, stable settlement. If agents make anything the native money of machines, it will not be Bitcoin — it will be a stablecoin on a fast rail, and Bitcoin's role will be that of a collateral asset, not a currency. This is not a bearish call on price. It is a structural call on function, and the two have been decoupled since the ETF approved.
If you want to watch this thesis resolve in real time, watch on-chain, not on stage. Agent-originated transaction counts, stablecoin transfer velocity from programmatic addresses, gas-auction behavior during agent bursts, and the growth of identity-registry writes — these are the vectors that will tell you whether the machine economy is real before any earnings call does. Volume without conviction is just noise, but volume from machines that never sleep is a different kind of signal entirely.
Now the counter-intuitive angle, because the consensus crypto take on agents is wrong in a specific and expensive way.
The consensus says agents will make crypto the native economy of machines, and therefore crypto wins the interface. I think that is backwards. The interface layer — the point of intent, the place where humans state goals — will be captured by Big Tech, because Big Tech owns distribution and identity at the consumer and enterprise level. Google owns the browser and the phone. Microsoft owns the desktop and the office. Apple owns the device. None of them will hand the entry point to a permissionless protocol, and none of them need to. What crypto wins is not the interface. What crypto wins is the rail beneath it — the settlement, identity, and verifiability layers that centralized stacks cannot provide without becoming the very intermediaries they claim to be disintermediating.
The decoupling thesis is this: agents will accelerate the split between the interface and the settlement layer, and crypto's prize is the second, not the first. Every agent that executes a task on behalf of a human creates a verification problem — did it do what it claimed, with what authority, at what cost? Centralized stacks can answer that with a log file only until they cannot, and the moment the answer carries financial or legal weight, the log file becomes an affidavit, and an affidavit requires a neutral substrate. That is a settlement-layer function. Crypto does not need to win the interface to win the decade. It needs to be the place where the interface's claims are settled, and that is a position Big Tech is structurally unwilling to occupy, because occupying it means giving up the intermediary margin.
The trap, as always, is timing. The floor is a trap for the impatient. The infrastructure thesis is correct and early, which are the two conditions under which most capital loses money by being right too soon. The protocol war will fragment before it consolidates. The first agent-security incident will trigger a regulatory response that lands hardest on the permissionless layer. And the token that "should" win will spend eighteen months doing nothing while the narrative churns.
So position for the rail, not the interface, and position slowly. Watch three signals and nothing else: the end-to-end success rate of agents on forty-step tasks, because that number, not any benchmark, decides whether the economy is real; the adoption curve of MCP versus A2A, because protocol coalitions resolve faster than protocol architectures; and the first machine-speed coordination failure in a DeFi fee market, because it will tell you which settlement layers are actually load-bearing. In a sideways tape, the edge is not conviction about direction — it is patience about structure. The agents are coming. The question is not whether you will catch the bottom of the narrative. It is whether you will still be holding the rail when the interface changes hands.