Agents Can't Buy Happiness. But…

A figure made the rounds early this week without really causing a scandal, and perhaps that is what ought to scandalise us. Nvidia, a graphics-chip manufacturer turned dominant supplier of the AI revolution, now carries a market capitalisation of around 5,460 billion dollars. The healthcare sector of the S&P 500 — fifty-nine companies, including Johnson & Johnson, Eli Lilly, Pfizer, Merck and UnitedHealth — totals roughly 5,200 billion. A single company is worth more than the listed industry most strategic to the ageing of the developed world. The four American hyperscalers alone have announced more than 725 billion in AI infrastructure spending for 2026, a major share of which feeds, directly or indirectly, the Nvidia ecosystem. The financial gravity of the world has a centre, and it is made of silicon.

That gravity creates its own field. Around it, two distinct orbits have organised themselves over the past eighteen months. The first, that of the foundation-model laboratories: on 31 March, OpenAI announced the close of a 122-billion funding round at a post-money valuation of 852 billion; Anthropic has been valued at 380 billion since its mid-February Series G; xAI exceeds 200. The second, more recent and more diffuse: the nebula of so-called "agent" companies — Sierra at 15.8 billion, Cursor in talks at more than 50 billion, Harvey at 11 billion since March, Cognition AI at 10. According to several sector surveys, AI captured around 80% of global venture capital in the first quarter of 2026, in the order of 240 billion dollars. Four of the five largest funding rounds in the history of private companies were closed in three months. All these valuations are reported as of mid-May 2026, in a sector whose volatility is priced in.

The magnitude of the figures says nothing about their soundness. It only says where American finance thinks value is concentrating. And that, it seems to me, is precisely where one should pause to look.

The word "agent" today designates a category of software capable of carrying out tasks in several steps, without constant supervision: an unhappy customer to call back, an invoice to dispute, a case file to prepare, a piece of code to fix. The narrative carrying it is straightforward. The chatbots of 2023 answered; the agents of 2026 execute. The market on offer is titanic — Bret Taylor, founder of Sierra and chairman of OpenAI's board, puts global customer service alone at 400 billion dollars a year, most of which would shift to the agent. By that yardstick, a 15-billion valuation for a three-year-old startup seems almost modest.

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

Eighteen months ago — and I invite anyone to verify in the press archives of autumn 2024 — a flourishing economy was organising itself around another tutelary category: the prompt. Prompt marketplaces raised funds. Prompt engineering schools opened their doors. Consultants sold libraries of professional prompts for fiduciary firms, law firms and HR departments at CHF 800 a day. The prompt appeared, at the time, as a new, defensible, monetisable skill. It had its lexicon, its certifications, its gurus. The promise was that whoever mastered the prompt would gain the upper hand over whoever merely used the AI.

Eighteen months on, almost all of that economy has melted away. Not because prompts are useless — they still serve a purpose — but because the value one claimed to enclose within them did not keep its defensive promise. The models absorbed what had to be absorbed. They now accept instructions in ordinary natural language, reformulate ambiguous queries, and produce from a curt request what one had to extract by forceps the day before. What was a skill has become a commodity. The marketplaces have closed. The schools have pivoted. The gurus have changed their watchword.

In large numbers, they have chosen the new one: the agent.

In my view — and I say this as an operator, not as a commentator — the agent is not a technological category. It is a figure of speech, designating the envelope through which a language model is entrusted with access to several tools, with a session memory, and with the ability to chain multiple calls to those tools before returning an answer to the user. Three elements, then: an orchestration loop, a set of plugged-in tools, a little memory. In proprietary lines of code, the whole weighs less than a modern accounting middleware.

And, above all, the whole is being absorbed by the very platforms that produce the models. In recent months, the large laboratories have begun publishing a banal but powerful mechanism: a folder of a few files — a procedure, sometimes a script, sometimes nothing more than rules of voice — that the model loads on the fly when it recognises the need. The function covers everything one called yesterday a legal agent, a fiduciary agent, a customer service agent, with the difference that no orchestration startup intervenes in the chain. The layer that the agent startups claimed to build and defend is being commoditised by the very laboratories that produce the models.

This absorption is neither an accident nor a low blow. It is the mechanical consequence of an economy in which the general intelligence of the model progresses faster than the abstractions built on top of it. An agent that made sense in September 2024 — when one had to hard-code task decomposition, tool selection and error handling — loses its raison d'être when, twelve months later, the model does all of that from two paragraphs of instruction. The agent is not dead. It has become a flat folder of instructions, slipped to a generalist model.

One must, here, distinguish two registers of value that the financial press conflates. The agent layer has, in my view, a real operating value — access rights, audit logs, connectors to legacy systems, human escalation handling, testing, service level agreements, regulatory compliance. A large enterprise putting AI into production needs all of that. The agent layer is, at that level of requirement, a legitimate product. What it does not necessarily have is technological rent value — the kind that would justify ten or fifty billion in capitalisation. The first can be absorbed tomorrow by native integration into the existing productivity or CRM platforms. The second would presuppose a defensible technical asymmetry that the evolution of the models erodes month after month. The whole of this note hinges on this: the agent layer has operating value; in its generalist version, it has no technological rent value. The market is paying today for the second. It will end up paying only for the first.

The consequence is uncomfortable for many recent funding rounds. If rent value fades, where, all things being equal, will durable value concentrate?

Upstream, first, where it is already concentrating: on the chip, the model, the data centre. Nvidia, OpenAI, Anthropic, xAI, Google. What the stock market is paying for, in weighing Nvidia more heavily than the entire listed healthcare sector of the United States, is precisely that conviction. That the value of the decade lies in the production of raw intelligence, not in the layer placed on top of it.

Downstream, then, where it is proved at the point of contact with the user: in proprietary data, vertical know-how, contractual trust, the regulatory chain. A Valais doctor holds patient data that no agent startup will ever reproduce. A fiduciary firm in Sierre holds cantonal tax and wealth knowledge that ten years of product development would not suffice to acquire. A winegrower in Salquenen holds a grape variety, a cadastre, a commune, a clientele. That value is not absorbable by an orchestration layer. It is the inverse of the agent: singular, situated, legally rooted, economically defensible.

To capture a raw skill where it is abundant, to bring it to the precise trade that needs it, to root it in a practice that retains responsibility for it. The logic of the bisse, again.

The title of this note poses an implicit question. If the agent does not bring happiness — in the sense that the rent value the financial centre is projecting onto it will probably not hold the decade — then what?

Then this, in my view. The underlying mechanism is in fact living up to its promises. Not in the dramatic version announced by the funding rounds, but in the practical version that a cantonal operator experiences daily. A four-person practice equipped with a generalist model and a folder of professional instructions — without any intermediate agent, without any third-party platform, without any subscription to Sierra or Harvey — handles its cases at a factor of four to five in productivity. A valley doctor keeps his patient records on Swiss rather than American infrastructure, and gains in professional trust what the agent claims to offer in automation. An alp cooperative documents its production with a folder of a few files that no agent valued at 50 billion could match in local relevance.

The agent does not bring happiness. But the work it was supposed to handle does. Provided it is entrusted to the right tool, at the right scale, and at the right remove from the non-European orchestration layers that practical sovereignty keeps at arm's length.

I do not know, at the time of writing, at what rhythm the absorption will take place. The financial narrative that values Sierra at 15.8 billion may hold for six months, eighteen months, three years. Bret Taylor himself — founder of Sierra, chairman of OpenAI's board, one of the architects of the wave — is announcing a correction. He speaks of an excitement so authentic, he says, that it is drawing in too much capital and too many companies; he predicts a culling effect, a consolidation in which almost all the actors will disappear. For lack of a better word, one will call that a shake-out. And it will have, on the cohort of agents, the effect that the correction of 2024-2025 had on the cohort of prompts: a great silence, and a few residual brand names that technological historiography will mention in a footnote.

In this space, Valais has an opportunity that should not be confused with a fad. The mistake would be to imitate the financial centre by betting on the agent — a cantonal administrative agent, a "Valais" legal agent, an alpine tourism agent. The right opportunity consists in betting on what the stock market, paradoxically, is already financing: the availability of a powerful general intelligence, brought to the trades of the canton without passing through the orchestration layer that claims to sublimate it.

The bisse does not manufacture water. It moves it. The agent will not manufacture intelligence. It claimed to wrap it. The market will eventually notice, as it noticed for the prompt, that one does not wrap a commodity.