The Agent Won't Buy You Happiness. But…

Nvidia is now worth more on the stock market than all fifty-nine healthcare companies in the S&P 500 combined. The financial cathedral rising around artificial intelligence has a name carved above its door: the agent. It will, I suspect, meet the fate of the prompt eighteen months ago.

A figure made the rounds early this week without causing much of a stir, and that alone should give us pause. Nvidia, once a maker of graphics chips and now the dominant supplier to the AI revolution, is worth around 5,460 billion dollars on the stock market. The healthcare sector of the S&P 500 (fifty-nine companies, including Johnson & Johnson, Eli Lilly, Pfizer, Merck, UnitedHealth) totals about 5,200 billion. A single company is worth more than the most strategically important listed industry for the ageing societies of the developed world. The four American hyperscalers alone are announcing more than 725 billion in AI infrastructure spending for 2026, a large share of which feeds, directly or indirectly, into the Nvidia ecosystem. The financial centre of gravity of the world has a location, and it sits in silicon.

That gravity generates its own field. Around it, over the past eighteen months, two distinct orbits have formed. The first belongs to the foundation-model labs: OpenAI announced on March 31 the close of a 122 billion round at an 852 billion post-money valuation, Anthropic has been valued at 380 billion since its Series G in mid-February, xAI is above 200. The second, newer and more diffuse, is the swarm of so-called agent companies: Sierra at 15.8 billion, Cursor in talks above 50 billion, Harvey at 11 billion since March, Cognition AI at 10 billion. According to several industry trackers, AI captured around 80 percent of global venture capital in the first quarter of 2026, on the order of 240 billion dollars. Four of the five largest funding rounds in the history of private companies closed within three months of each other. All these valuations are dated to mid-May 2026, in a sector whose volatility is already baked into the price.

The scale of these figures says nothing about whether they're justified. It only shows where American finance believes value is concentrating. And that, it seems to me, is exactly where we should stop and look closely.

The word "agent" now designates a category of software capable of carrying out multi-step tasks without constant supervision: calling back a dissatisfied customer, disputing an invoice, handling a case file, fixing a piece of code. The story behind it is disarmingly clean. Chatbots in 2023 answered; agents in 2026 act. The market on offer is vast: Bret Taylor, founder of Sierra and chairman of OpenAI's board, puts the value of global customer service alone at 400 billion dollars a year, most of which is expected to shift to agents. Measured against that, a 15 billion valuation for a three-year-old startup starts to look almost modest.

This is where the short memory of markets and the long memory of practitioners part ways.

Eighteen months ago (and I'd invite anyone to check the press archives from the autumn of 2024), a thriving economy had organised itself around another tutelary category: the prompt. Prompt marketplaces raised funding, schools of prompt engineering opened their doors, and consultants sold libraries of trade-specific prompts, for fiduciary firms, law offices, HR departments, at eight hundred francs a day. The prompt looked, at the time, like a new skill: defensible, monetisable. It had its own vocabulary, its certifications, its gurus. The promise was that whoever mastered the prompt would hold the upper hand over whoever merely used AI.

Eighteen months later, almost that entire economy has evaporated. Not that prompts are useless, they still serve a purpose, but the value people claimed to have locked inside them never delivered on its defensive promise. The models absorbed what needed absorbing. They now accept plain, everyday instructions, rephrase ambiguous requests on their own, and produce from a terse ask what used to require painstaking extraction. What was a skill became a commodity. The marketplaces closed, the schools pivoted, the gurus changed their tune.

Many of them, in large numbers, picked the new one: the agent.

I say this as an operator rather than a commentator: the agent has less to do with a technological category than with a figure of speech. The word describes the wrapper by which a language model is granted access to several tools, a session memory, and the ability to chain multiple calls to those tools before handing a response back to the user. Three components, then: an orchestration loop, a set of connected tools, a bit of memory. The whole thing amounts, in lines of proprietary code, to less than a modern accounting middleware package.

And, more to the point, the whole thing is now being absorbed by the very platforms that produce the models. In recent months, the major labs have started rolling out a mechanism that is both mundane and powerful: a folder of a few files (an operating procedure, sometimes a script, sometimes nothing more than voice-and-tone rules) that the model loads on the fly when it recognises the need. This function covers everything that used to be called a legal agent, a fiduciary agent, a customer-service agent, with one difference: no orchestration startup sits in the chain anymore. The layer that agent startups claimed to be building and defending is being commodified by the very labs that produce the models themselves.

This absorption is neither an accident nor a low blow. It's the mechanical consequence of an economy in which a model's general intelligence advances faster than the abstractions layered on top of it. An agent that made sense in September 2024, when task decomposition, tool selection, and error handling all had to be hard-coded, loses its reason for being once the model, twelve months on, can do all of that from two paragraphs of instruction. The agent survives, but reduced to a flat instruction folder handed to a general-purpose model.

Here it helps to separate two registers of value that the financial press keeps blurring together. The agent layer has real operational value: access rights, audit logs, connectors to legacy systems, management of human escalations, testing, service-level agreements, regulatory compliance. Any large enterprise putting AI into production needs all of that, and at that level of rigour the agent layer is a legitimate product. The rent-seeking technological value, though, the kind that would justify ten or fifty billion in market capitalisation, is a different matter: it would require a defensible technical asymmetry that the evolution of the models erodes month after month, whereas the operational value can be absorbed tomorrow by native integration into existing productivity or CRM platforms. That's the whole story in a nutshell. The market is paying for the rent today. Eventually it will pay only for the operations.

This is an uncomfortable conclusion for a great many recent funding rounds. If the rent value fades, where does lasting value end up concentrating?

Upstream, first, where it's already concentrating: in the chip, the model, the data centre. Nvidia, OpenAI, Anthropic, xAI, Google. What the stock market is paying for, by weighing Nvidia heavier than all of listed American healthcare, is precisely this conviction: that the decade's value lies in producing raw intelligence rather than in the layer built on top of it.

Downstream, second, where it proves itself in contact with the user: in proprietary data, professional know-how, contractual trust, the regulatory chain. A doctor in Valais holds patient data that no agent startup will ever reproduce. A fiduciary firm holds cantonal tax and wealth-management knowledge that ten years of product development wouldn't be enough to acquire. And a winegrower, a grape variety, a land registry, a commune, a client. That kind of value escapes every orchestration layer, because it is the exact opposite of the agent: singular, situated, legally rooted, economically defensible.

Capture a raw capability where it's abundant, carry it to the specific trade that needs it, root it in a practice that keeps responsibility for it. The logic of the bisse, once again.

The title of this piece poses an implicit question. If the agent won't buy you happiness, in the sense that the rent value the financial world is projecting onto it likely won't survive the decade, then what?

Then this: the underlying mechanics are, in fact, delivering on their promise. Less in the dramatic version announced by funding rounds than in the practical version that a cantonal operator experiences every day. A four-person firm equipped with a general-purpose model and a folder of trade-specific instructions (no intermediary agent, no third-party platform, no subscription to Sierra or Harvey) processes its cases at four to five times the productivity. A doctor in an alpine valley keeps patient records on Swiss rather than American infrastructure, and gains in professional trust exactly what the agent claims to offer through automation. An alpage cooperative, meanwhile, documents its production with a folder of a few files that no agent valued at 50 billion could match for local relevance.

The agent won't buy you happiness. But the work people claimed to be handing over to it does. Provided it's handed to the right tool, at the right scale, and at the right distance from the non-European orchestration layers that practical sovereignty keeps at arm's length.

As I write this, I don't know how quickly the absorption will unfold. The financial story that values Sierra at 15.8 billion might hold for six months, eighteen months, three years. Bret Taylor himself (founder of Sierra, chairman of OpenAI's board, one of the architects of this wave) is forecasting a correction. He speaks of an excitement so genuine, in his words, that it's drawing in too much capital and too many companies; he predicts a culling effect, a consolidation in which nearly every player disappears. Call it a clearing-out. It will do to the agent cohort what the 2024–2025 correction did to the prompt cohort: a great silence, and a handful of residual brands that the history of technology will mention in a footnote.

Valais has, in this space, an opportunity that shouldn't be mistaken for a trend. The mistake would be to imitate the financial world by betting on the agent: a cantonal administrative agent, a "Valais" legal agent, an alpine tourism agent. The right opportunity is to bet on what the stock market, paradoxically, is already financing, namely the availability of a powerful general intelligence, carried all the way to the canton's trades without passing through the orchestration layer that claims to elevate it. It is, once more, the gesture of the bisse: moving water it did not make. The agent, for its part, claimed to wrap that water in something more. The market will eventually notice, as it did with the prompt, that you can't wrap a commodity in anything.