Google Turned Search Into an Agent. The Accountability Layer Is Still a Ghost.

$180 billion. Maybe $190 billion. That is Alphabet's stated annual capital expenditure, roughly six times the $31 billion it booked in 2022. I read that figure three times before I understood what it was actually telling me. It is not a spending flex. It is an admission — that the search box is no longer a place you type into. It is a background process that runs without you. And background processes cost money in a way that a query spike never did. The framing at I/O was polite: information agents, agentic booking, a generative UI that assembles itself per query. The engineering reality underneath is uglier and more interesting. Search is migrating from a request-response architecture to a resident, stateful agent architecture. That migration is the story. Everything else is keynote gloss. Here is what actually shipped, stripped of the marketing. Gemini 3.5 Flash becomes the default model — not because it is the most capable, but because it is four times faster and less than half the price. Information agents are defined as persistent, 24/7 background monitors that reason across web pages, blogs, social posts, and live data feeds, with the ability to take action. Agentic Booking rolls out to US users. Generative UI renders interface code on the fly. And Antigravity graduates from a coding environment into a platform for developing and governing fleets of autonomous agents. I have watched this movie before. In 2020, during DeFi Summer, I wrote a Python script that monitored Uniswap and SushiSwap pools for arbitrage. It executed over 500 trades and cleared $45,000. The lesson was not that I was clever. The lesson was that the incentive structure — not the ideology — moved the market. Yield farming was never about decentralization. It was about who captured the emissions. Arbitrage is just geometry disguised as finance, and the geometry always points at the incentive. Google just moved its own incentive structure. The question every crypto operator should be asking is not whether this is good for AI. It is what happens to the flow of value when the entry point becomes the executor. When the thing that used to hand you a link starts spending your money instead. Let me define the three new load types, because the compute bill only makes sense through them. One: background resident monitoring. Continuous, low-frequency, long-duration. This raises the baseline power and compute draw. It is not a training spike you can schedule around. It is a floor, and floors do not go away at night. Two: generative UI and Antigravity workloads. Bursty, compute-heavy, stateful. Peak load stacked on top of that floor, plus the persistence cost of remembering what every session built. Three: voice agents calling merchants. Real-time, low-latency streaming ASR/TTS plus an LLM, which demands regional or edge deployment. This is the most fragile piece and the most likely to be overvalued. It touches dual-consent recording laws, human-versus-bot disclosure duties, IVR navigation, dialects, and retries. There is no API for a plumber. There is only a phone number and a human being having a bad day. That last point is where I part ways with most of the coverage. Everyone is modeling the agent as a software problem. The phone channel is a legal and reliability problem wearing a software costume. One failed booking is not a slightly wrong paragraph. It is a monetizable loss, and someone has to eat it. Now the model choice, because the model choice is the tell. Flash as the default is Google competing on cost, not capability. That is a confession that the model layer has commoditized. If the frontier is no longer differentiable, you cannot price on intelligence. You price on distribution. And Google's distribution is not a model. It is search plus Chrome plus Android — a triple default no competitor can assemble. I spent three months in 2024 reading ETF prospectuses, and I learned something about language: when an issuer leads with a price-and-speed claim, they are defending a moat, not expanding one. The capability claims come later, and they come thin. Here they came with zero benchmarks, zero reliability metrics, zero error rates on long-horizon tasks. Which raises the question nobody at I/O answered. What is the success rate of a ten-step agent task? If each step is 95% reliable, the chain survives at roughly 60%. Five steps and you are already losing money on bookings. The reliability math for agentic commerce is unforgiving, and it is worse than summarization — because in summarization a stumble is a bad sentence, while in booking it is a charge on a card. The downstream consequence deserves its own paragraph, because it is where content and crypto overlap. Search changing from a traffic distributor into a demand terminator cuts the value chain that funds open content: content, then search, then click, then ad or conversion. Raise the zero-click rate and you starve the source. Cut the income and production motivation falls. Production falls and public content supply shrinks. Shrink the supply and the agent's own retrieval and training sources degrade. It is a self-weakening loop, and Google is inside it, not above it. If clicks vanish, Google's ad machine has to be rebuilt in the same breath. There is a plausible window where usage rises while ad revenue sags. Then the accounting layer. Capex of $180–190 billion against $31 billion in 2022 is a six-fold jump. It implies the eighth-generation TPU plus owned data centers are the real strategic asset — the attempt to vertically integrate chip, model, and distribution to crush unit cost against competitors who rent from Nvidia. But capex like that lands on the income statement as depreciation and on the grid as baseline load. Watch Cloud margin and depreciation absorption, not the demo reel. And the usage numbers — 2.5 billion AI Overviews users, 1 billion AI Mode monthly actives — are not revenue. They are exposure. A passive AI Overview and a deliberate agent task are both counted, and they are not the same behavior. Worse, queries at an all-time high has a perverse explanation: users trying to get past the AI summary rewrite their query over and over. That is query inflation, not demand growth. I do not trust the narrative. I read the code, and here there is no code to read — only self-reported exposure. Here is the angle almost nobody is running. The most accurate sentence in the entire discourse is that the governance stack is missing. But that is not a technology gap. That is an accountability gap. And it is the single largest unclaimed opportunity in the agent economy — which makes it a crypto problem whether or not crypto shows up. Walk the chain. If an agent books a flight and it is wrong, who pays? Google, the merchant, or the user? If a background monitor aggregates data across a dozen sources and leaks, what is the legal basis for that combination under GDPR? If a voice agent impersonates a consumer to a hairstylist, whose disclosure obligation is it? These are not inference problems. They are insurance, contract, and identity problems. They will not be solved by a faster model, because a faster model does not decide who is liable. This is precisely where I expected crypto to be building. And it is not. In 2026 I built a prototype where an autonomous agent negotiated data-access fees over Ethereum, managing a wallet with $10,000 in testnet funds. The hard part was never the payment. It was authorization — who the agent was acting for, what the scope was, and how to revoke it mid-task. That is verifiable credentials, delegation semantics, and on-chain attestation. It is the exact toolset crypto keeps inventing for other purposes and then abandoning. Instead, the industry is launching another Layer 2 and another liquidity pool, slicing a scarce user base into fragments and calling it scaling. If the agent economy needs one thing, it is an authorization and dispute layer. Smart-contract escrow for agent actions. Revocation primitives. Accountability receipts. The wrong layer is being built, and it is being built loudly, which is how you know nobody is measuring the load. There is also a regulatory tripwire the coverage ignored. The EU's DMA already forces search choice screens. If a default agent's behavior is classified as gatekeeping, that default gets decoupled by law. Antitrust enforcement after the US search-monopoly ruling can constrain the whole strategy. That is a nearer and harder constraint than any AI ethics debate, and it sits directly on the moat I just described. The question is not whether search becomes an agent. That already happened, and the capex says Google committed to it before the keynote did. The question is who captures the accountability layer when a machine spends your money and gets it wrong. If crypto shows up with revocation, delegation, and dispute settlement, it earns a seat at the table. If it keeps shipping L2 number forty-seven, it watches the largest trust market in a generation get built in a language it does not speak.