Home / Explainers

Explainers

The Agentic Economy

Anatomy of commerce without buyers and of cognition delegated to machines

Danilo Giudice · February 18, 2026

Type

Explainer: from zero, with numbers and sources

Time

22 min read

Level

Intermediate

What you’ll get

• What changes when the web is designed for agents
• How x402’s pay-per-request payments work
• Where value moves when compute costs almost nothing

This article was first published in Italian on Medium on February 18, 2026.

Abstract

Months ago, I changed a website’s frontend so an AI agent could navigate it without getting stuck. That episode made me think about an emerging economy driven by AI agents. It could redefine market dynamics and, beyond them, the entire social, epistemological, and cultural fabric of today’s society. In this paper I want to explore the systemic implications of a shift from a web built around user experience (UX) to an infrastructure optimized for machine experience (MX). I’ll look at how this transition reshapes power relations, knowledge formation, individual identity, and governance structures. If interfaces start optimizing for software agents instead of human beings, what happens to the architecture of the web itself? When agents become the primary economic actors, human attention (the scarce resource the whole digital economy was built on) stops being the core of value extraction. In this scenario, the agentic economy is an anthropological paradigm shift. It forces a radical rethink of the categories we use to understand work, agency, freedom, and social control. And it raises an even more radical question. In this regime, the dominant economic variable becomes the marginal cost of compute, measured in tokens consumed, latency tolerated, and verifiable guarantees. What happens to the very idea of work, freedom, and social control?

0. Backstory

Months ago I was doing my job as an onchain data analyst, juggling raw data, .csv files, and charts to write yet another research report. That same week, OpenAI released its agent. It grabbed my attention right away, and I figured it could help me with the tedious job of collecting that data. I gave it meticulous instructions over several prompts. Then I sat back and watched it work its way through the pages of my company’s tool. It moved between domains and handled buttons and tooltips quite naturally. It was painfully slow. Still, I figured that even at that speed, working in the background, it could save me several hours. Then, all of a sudden, it got stuck. It couldn’t get past a dropdown menu, and it failed miserably at the task I had given it, one step from the finish line.

I contacted my company’s frontend team and asked for a small iteration: a small change to the interface of the tool where the agent was struggling. The team thought it over, but since it was a non-invasive change, they approved it. The next day the fix was committed.

I had the agent run the task again and… EUREKA! It completed the task and gave me the files I needed.

Machine experience

A few days later I realized that the change we had made was, in fact, our tool’s first change for a machine interface, not a human one. I admit it may sound trivial. But I had spent months with the UX/UI team studying microcopy and mapping the user journey. I even bought psychology books to improve the work we were doing. And all of a sudden we had committed an iteration for “machine experience” instead of “user experience.”

That realization is why I wrote this article: I wanted answers to questions that had been troubling me for months.

1. Introduction and thesis

Shoshana Zuboff has documented how surveillance capitalism turned human experience into raw material for extracting predictive value. Today’s digital economy is built on capturing and monetizing attention, with interfaces designed to maximize engagement, time on site, and behavioral dependence. This is also part of what I do for a living, which is why it struck me so hard.

The agentic economy turns this logic upside down: when software agents become the main economic actors, human attention is no longer the scarce resource that value extraction depends on. Agents have no attention span to capture. They don’t fall for dark patterns. They aren’t vulnerable to persuasion techniques based on cognitive biases, or to user journeys carefully mapped and tuned to drive conversions. They run on deterministic policies and optimize for efficiency and marginal cost, so they are immune to the seduction of design.

This creates a fascinating paradox. On one hand, humans are “freed” from having to pay constant attention to choosing and transacting. On the other, their ability to influence the market through preferences expressed in real time withers away. Agency shifts from direct choice to setting policies and constraints up front (for those who already know the topic, that’s the agent’s rules.md). The individual goes from active consumer to policy maker of their own delegation.

But here comes the first contradiction. If Herbert Simon was right that information consumes the attention needed to process it, what happens when attention is delegated entirely? The knowledge behind economic decisions ends up in a layer humans can’t reach, encoded in parameters and constraints that the agent acts on but the individual can no longer experience directly.

Today’s digital economy rests on a specific ontology: the human user as the central node of the economic act. The user browses, compares, gets persuaded, buys, and, above all, leaves behavioral traces that can be monetized. This economic subject is both consumer and product: it consumes goods and services, but it produces data that feeds profiling architectures and big data. Yet this paradigm, so established that it seems natural, is already wobbling. You can see some signs of it today. GEO and AIO (optimizing content so that LLMs index it) are already undermining this logic. Take Gemini, now built right into Google’s search engine. The user no longer gets a list of results to choose from on their own, browsing several sources and building their own synthesis. Instead, they get funneled toward answers generated directly by the model. The sources become footnotes, supporting citations for a narrative the AI has already assembled. So the human goes from active explorer to passive validator, who can no longer synthesize information and only consumes the prepackaged summaries generated for them.

On the cognitive side, this scenario fits the idea that in information-rich environments the scarce resource is attention, and that information “consumes” attention. That makes it rational to delegate filters and repetitive decisions.

This is exactly the fertile ground for an agentic economy, where tasks and agency are delegated to personal agents.

2. Model: UX Web vs. MX Web

We can call today’s web the “UX Web”: the web where humans interact with a visual interface and the primary unit is the session. A human visits a website, explores options, interprets content, makes decisions, accepts terms, pays, and then handles exceptions (delays, returns, disputes). This interface (UX/UI/frontend) is how discovery, persuasion, consent, and payment happen. It is also a highly specialized professional field, full of people who study, spot, and apply methods that improve this experience and, in some cases, standardize it.

Now imagine the opposite scenario, the “MX Web.” Here the primary unit becomes a pipeline that can be standardized (that is, automated) and that an agent can run at scale:

1. Discovery: find counterparties/capabilities (not “pages”).
2. Quote/terms: get explicit, machine-readable conditions (price, SLA, limits, policy).
3. Payment: settle pay-as-you-go, ideally per request.
4. Execution: call the capability and receive the output.
5. Proof/verification: measure or validate the output where the risk calls for it.
6. Reputation update: update reliability signals (for future choices).

This pipeline is the “industrial” equivalent of human browsing. It optimizes for determinism, throughput, and audit, not for engagement.

So it’s fascinating to imagine a scenario where UX falls away, and with it all the related best practices: GUI, user journey, copywriting, marketing, etc. In the resulting scenario, agents will talk to each other in machine language, trading raw data with no container a human can make sense of.

Research is tedious: comparing flights across dozens of aggregators, finding coffee capsules at the best price, tracking down that one piece of clothing in pure mackinaw wool. All of this takes hours, if not days. Some people love it (I do, for one). But it’s just as true that an agent doing these chores for you would save a huge amount of time, and you could spend it on something else.

The human would be left with governance: policies, authorizations, and audit/override when something goes off track. The scenario taking shape is an ecosystem of agents and merchant endpoints, where the economic act changes from “I visit a site and choose” to “I set constraints and authorize an agent to negotiate/execute.” This tears down the usual infrastructure. The frontend becomes unnecessary, and agents would move through MCP (Model Context Protocol) servers.

I admit the last two paragraphs are a stretch. The frontend wouldn’t disappear completely: humans will always need an interface their own “species” can understand, to settle disputes, run audits, or deal with missed payments. The provocation still stands, though. It rests on how this systemic delegation of research and execution to autonomous agents will reshape the World Wide Web as we know it today.

3. The great delegation: from the division of labor to the division of agency

Karl Polanyi described the “great transformation” as the process by which the market split off from society and subordinated social relations to the logic of exchange. The agentic economy could be a second great transformation: the separation of economic agency from the individual.

In Adam Smith’s division of labor, individuals specialize their skills but keep the ability to choose, bargain, and evaluate. In the agentic division, individuals delegate more than execution: they delegate the very capacity for informed choice. It’s no longer “I buy X because I evaluated it” but “my agent bought X within the constraints I gave it.”

This systemic delegation reshapes moral and economic responsibility. If an agent negotiates a bad contract, whose fault is it? The individual who set the constraints badly? The developer who trained the model badly? The provider who sold an unreliable capability? Or the human who copy-pasted a rules.md they found online? The chain of accountability shatters into a web of mutual delegations, where finding the locus of responsibility becomes an exercise in interpretation (and a headache for law firms).

Herbert Simon talked about “bounded rationality”: humans make “good enough” decisions (satisficing) because computing the optimum costs too much. Agents remove this constraint. They can explore huge spaces of possibilities and converge on optimal solutions for defined metrics, something the average human can’t do (lol). But this creates a new problem: individuals no longer have any way to understand the choices made in their name. They can only check, after the fact, that those choices match constraints they may not even remember setting. What psychological effects could all this have?

4. The inverted Panopticon

Jeremy Bentham imagined the Panopticon as an architecture of social control built on the possibility of being watched at any moment. Michel Foucault turned it into the metaphor of the disciplinary society: power no longer needs to punish; it is enough that it can see. In his novel 1984, Orwell probably merged the two.

What changes in an agentic economy is the arrival of a kind of inverted surveillance. Instead of a central authority watching the many, the many (agents) watch and certify every interaction, in a distributed way. And since every human can have countless agents, built and configured for different purposes and with different skills, the agents end up outnumbering the humans by monstrous orders of magnitude.

In crypto, for example, innovation moves faster (probably also because there’s less regulation). Standards like ERC-8004 are designed to contain the accountability problem and also to pin down an agent’s identity. They create permanent reputation registries, where every transaction leaves a trace, every breach is recorded, and every anomalous behavior is flagged.

The result is a distributed Panopticon where reputation becomes the new form of social control. Central authorities handing out sanctions are no longer needed: if your Reputation Score drops below a threshold, agents simply stop doing business with you (China’s social credit system: sound familiar?). Governance becomes algorithmic, and behavior is shaped by incentives coded into protocols, instead of by internalized social norms. It’s game theory pushed to the extreme, where the main incentives are the ones that govern how machines act, not human drives or feelings.

But who watches the watchers? If reputation is onchain and immutable, how do you handle the right to be forgotten? How do you tell an honest mistake from malicious behavior? How do you stop systemic biases (for example, algorithmic discrimination) from hardening into permanent exclusion?

Thomas Hobbes argued that individuals give up freedom to a centralized Leviathan in exchange for security. In the agentic economy, the Leviathan dissolves into a thousand distributed micro-authorities (agents, protocols, registries). They have no formal sovereignty, but they exercise power through reputational exclusion. It’s an automated state of nature, where trust is built through cryptographic checks instead of social relationships.

5. The extinction of the interface: Marshall McLuhan and the disappearing medium

Marshall McLuhan said that “the medium is the message”: content isn’t the only thing that counts, because the form that carries it shapes perception and thought. The web interface (UX/UI) is more than a neutral container of information: it is the medium through which human beings learned to navigate digital knowledge. Fields like neuromarketing, applied psychology, microcopy, etc., have made a great many companies rich.

Yet the agentic economy would make this medium obsolete. Agents don’t need visual interfaces. They navigate through APIs, protocols, and MCP servers, where design falls away, copywriting becomes irrelevant, and the user journey dissolves into machine-readable pipelines.

But what happens when the medium disappears? McLuhan argued that media are extensions of man: the wheel extends the foot, the book extends the eye, the computer extends the central nervous system. The web interface was an extension of the human ability to explore and choose. When that extension is removed and replaced by autonomous agents, what withers in us?

The ability to compare, to evaluate, and to tell good quality from bad could become vestigial. Just as spatial memory withered with the arrival of GPS, the ability to navigate economic complexity could erode once we delegate it systematically.

Individuals become strangers to the decisions made in their name. They can only check, after the fact, that those decisions comply on paper, with no real claim to understanding them.

Is this the TikTok of the agentic era? (Go ahead, get pissed off.)

6. Small-change economics: pay-per-request as the infrastructure of granularity

So what would an agentic economy look like at a granular level? What would change, in macroeconomic terms, in the digital sector?

Today the digital economy is organized around artificially bundled units of consumption: monthly subscriptions, enterprise packages, service bundles, freemium with arbitrary gates, paywalls, etc. This structure is no accident. It responds to specific technical and economic constraints: the transaction cost of micropayments has historically been prohibitive. That made it rational to “package” value into sellable units that spread out the administrative, contractual, and financial overhead.

The result is an economy of forced overconsumption: we pay Spotify €9.99 a month even if we only listen to music for 40 minutes. We keep subscriptions to services we use now and then because “I’ve already paid for them anyway.” We pile up access to platforms and use a tiny fraction of their capabilities.

The pay-per-request model flips this logic, and it only works in an agentic economy that removes the need for accounts, contracts, conversion funnels, and retention newsletters. The economic unit becomes the single action: a database query, one inference on a model, retrieving a document, running a computation, or a single API call. And this atomization of payments becomes a revolution in economic composability.

For a software agent, pay-per-request means it can build arbitrarily complex pipelines out of different microservices, without negotiating a contract with each provider in advance. An agent can query five different data sources, review twelve articles, apply seven quality filters, buy the remaining content, summarize it with a language model, and check the output with a fact-checking system. All of this becomes a stack that runs automatically. Each step costs a fraction of a cent and needs no manual setup. And, once again, the crypto world got there first, alas.

6.1 x402: HTTP as a payment rail

The x402 protocol formalizes an elegant pattern that builds payment right into the HTTP flow:

1. Request: The agent sends a standard HTTP request to an endpoint
2. Quote: The server replies with 402 Payment Required, including the price, the terms, and the accepted payment method
3. Authorization: The agent repeats the request with a payment authorization attached (a cryptographic signature, an on-chain commitment, a state channel)
4. Delivery: The server verifies the payment and delivers the requested output

This choreography looks trivial, but it solves a problem that has paralyzed the web for decades. HTTP status code 402 has existed since 1997, but it went unused because there was no settlement layer efficient enough to make micropayments workable. Credit cards have fixed fees that make transactions under a dollar uneconomical. PayPal and similar systems require accounts, KYC (know-your-customer identity checks), and a link to a legal identity. First-generation blockchains had fees that were too high and throughput that was too low (I’d have a lot to say here, but I’ll keep quiet so I don’t look biased).

Combined with layer-2 networks, x402 removes this friction, so a transaction can cost a fraction of a cent in fees and settle in milliseconds. No sign-up, no mutual trust between the parties. The agent and the provider don’t even need to “know” each other in the traditional sense. They only need to speak the same protocol and accept the same settlement method.

6.2 Tokens as the atomic unit of value

For the most common chatbots (ChatGPT, Claude, Gemini, etc.), a token is roughly 3–4 characters of text (in languages written in the Latin alphabet). Tokens count both input (the prompt) and output (the answer). Commercial models are priced on exactly this metric: Claude Sonnet 4.5 costs $3 per million input tokens and $15 per million output tokens. This level of granularity turns compute into a perfectly divisible, composable commodity. An agent can use tokens from one provider, then switch to another based on price, latency, and reliability at that moment. In fact, routers and aggregators have appeared (Martian, OpenRouter, etc.) that act as a meta-layer. They receive a prompt and route the request to the cheapest, fastest, or most reliable provider, so the consumer never has to pick the final vendor.

But here a paradox appears: if tokens become a perfect commodity, how do you justify the price differences? GPT-4 costs 10x as much as GPT-3.5, and Claude Opus costs 5x as much as Haiku. The answer is output quality per token: not all tokens are equal. A token generated by a more capable model is “worth more” because it handles more complex tasks, makes fewer mistakes, and needs fewer retries (the so-called reasoning, context windows, etc.). Agents switch between the two on the fly depending on the task, to get the best cost/performance. Still, the slide toward commoditization is relentless: prices per token fell 90% in two years, and they will keep falling as specialized hardware arrives and competition grows.

6.3 Composability as a competitive advantage

The economic implications of this architecture run deep. In the current regime, integrating services takes partnerships, API keys, contracts, and often negotiated revenue sharing. This creates huge barriers to entry. Only big players can afford to integrate dozens of services, and only those with bargaining power can negotiate good terms. Anyone building a product today, even by vibe coding, knows what we’re talking about.

In the pay-per-request regime, composability becomes permissionless: any agent can call any public capability and pay only for what it uses. This democratizes access to sophisticated tech stacks. You don’t need to be Google to build pipelines that combine computer vision, NLP, knowledge graphs, and reasoning engines. You just need to be able to specify the sequence of operations and have the budget to run them (this one deserves an article of its own).

But this also creates new competitive dynamics. If everyone can access the same capabilities on a pay-as-you-go basis, where does value get captured? The answer has three parts:

1. Speed and latency: In markets where agents arbitrage inefficiencies in real time, a few milliseconds of advantage are worth orders of magnitude in profit. Whoever controls low-latency infrastructure (edge computing, optimized routing, preferential access to liquidity pools) extracts a rent from speed.

2. Reliability and SLAs: When you chain ten microservices in a pipeline, overall reliability is the product of the individual reliabilities. A service with 99.9% uptime looks solid, but chain ten of them and the probability of success drops to 99%. Whoever can guarantee 99.99% or 99.999% can charge a premium, especially in critical settings (finance, healthcare, infrastructure).

3. Verifiable guarantees: In a market of computing commodities, differentiation shifts to what you can prove. Can an AI inference service certify that the model hasn’t been backdoored? Can a data provider prove cryptographically that the data hasn’t been tampered with? Can an executor provide proof that it ran correctly? These guarantees are technically possible through zero-knowledge proofs, trusted execution environments, and distributed consensus, and they become the new competitive moat. (I’ve seen price quotes from big-name web3 auditors for hundreds of thousands of euros for a single smart contract. Ironically, that same smart contract got hacked months later, with losses in the millions.)

6.4 Converging on marginal cost

There’s a deep irony here. An economy driven by capitalist logic pushed to the extreme (maximum allocative efficiency, zero friction, perfect competition) converges on Marxian predictions: prices tend toward the marginal cost of production.

For compute, the marginal cost is mostly electricity and hardware depreciation (remind you of anything? Hint: Bitcoin mining). One LLM inference costs a fraction of a cent in energy. Distributed storage costs nanodollars per byte-second. Bandwidth is almost free. When markets are efficient and competition is global, margins collapse toward these physical floors. (If I remember correctly, the price of a barrel of oil also tends toward its marginal cost, but I found conflicting opinions online.)

You can already see this in cloud computing: compute and storage prices have been falling exponentially for twenty years. But so far, value capture has been protected by bundling, lock-in, and deliberate complexity. AWS sells compute, and with it an ecosystem that is expensive to leave. Google Cloud sells storage, and with it integration with Google’s advertising infrastructure. Azure sells IaaS, and with it compatibility with the Microsoft enterprise world.

Pay-per-request destroys these paradigms: if I can switch providers for every single request, based on price/latency/reliability at that moment, bargaining power flips. I’m no longer captive to an ecosystem; I can switch at no cost. Older Italian readers will remember the oligopoly of mobile carriers in Italy before Mobile Number Portability (MNP) went live in 2002. Portability brought a sharp drop in prices and the birth of new carriers.

But here comes a paradox: if margins collapse and value turns into a commodity, where does the profit go? It moves from capability providers to the controllers of the rails (the networks that payments and data travel on). Not those who run compute, but those who own the infrastructure for discovery, reputation, and settlement. Not those who sell storage, but those who certify reliability, guarantee SLAs, and insure against failure. At baccarat, the house always wins.

It’s the same dynamic we saw with platforms: Facebook doesn’t produce content, YouTube doesn’t make videos, Uber doesn’t own cars, and so on. They capture value by controlling the coordination layer. In the agentic economy, the coordination layer is made of protocols (x402, ERC-8004), reputation registries, dispute resolution systems, and providers of cryptographic guarantees.

6.5 Social implications

Pay-per-request radically changes the demand for intellectual work: what happens if everything can be consumed on demand? Expertise no longer lies in execution (gathering information, summarizing, formatting) but in specification: knowing which questions to ask, which constraints to set, how to validate output, and when not to trust automation.

It’s the same transformation we saw with the calculator: it didn’t eliminate mathematicians, but it made the skill of calculating by hand obsolete. The pay-per-request economy makes manual research, data entry, and routine processing obsolete. What remains is judgment, taste, and the ability to specify intention in a machine-executable way. To borrow from Rick Rubin, we won’t need talent anymore. We’ll focus on the refined taste of a few individuals. Writing the best lines of code stops mattering. Giving the right prompt, with only the result in mind, is what counts.

But this repeats the inequality we’ve seen with every earlier wave of automation. People with the cultural capital to specify sophisticated policies multiply their productivity. People who did routine work based on now-commoditized skills get pushed to the margins. A researcher who knows how to orchestrate agents can produce 100x more than before, and a data entry clerk gets replaced entirely. Looking at the present, we already have the first concrete example: devs vs. vibe coders.

And unlike earlier automation, this one is global and instant. You don’t need to open factories in Vietnam to exploit differences in labor costs. An agent in Italy can use compute in Singapore, data in Estonia, and validation in Switzerland, all in the same pipeline, which costs cents and finishes in seconds. Competition for intellectual work becomes truly planetary, with powerful deflationary effects on any activity that can be broken down into discrete tasks.

6.6 The risk of epistemic fragmentation

There’s one last aspect, subtler but potentially more destabilizing: when knowledge is consumed in micro-units (single queries, single papers, single fact-checks), you lose the context that comes from browsing holistically.

Reading a book is not the same as extracting 200 relevant quotes. Exploring a body of scientific literature is not the same as retrieving the 10 most-cited papers. Some knowledge comes from the process itself: from serendipity, from the unexpected connections you notice along the way. Claude Shannon would have recognized this as a problem of entropy: when you compress information by removing redundancy, you maximize efficiency but lose resilience. Will we still get writing as dense with information and cross-disciplinary thinking as Nassim Taleb’s?

Agents optimize for relevance to a specific query, without “wasting” time on detours, tangents, or curiosity that serves no direct purpose (especially if your rules.md has an economy rule about using tokens sparingly). But those detours are often exactly where breakthroughs, insights, and paradigm shifts come from (I recommend taking a look around LessWrong; it’s a wonderful example of an intellectual potpourri).

So there’s a risk of a local optima trap: settling on the best nearby solution and never finding a better one farther away. Agents get extremely efficient at optimizing for the goals they’re given, but they lose the ability to discover goals we didn’t know we should look for. Knowledge breaks up into a mosaic of micro-retrievals: highly targeted, but epistemically poor. That’s also because agents optimize to minimize the tokens they use (given what they cost), and this creates an economic disincentive to explore: the pressure is against waste.

We could imagine agents programmed to “explore” beyond what’s strictly needed, to follow tangents, to keep some serendipity. But that takes intent: economic efficiency alone will push toward maximum compression. Agents are too “efficient” to wander, get lost, and discover new things. Most likely, in the future we’ll reward cognitive traits we now tend to curb, like ADHD, with its ability to hold a linear focus combined with an incredible knack for tangential associations. It might be exactly what an economy of hyper-optimized agents can’t replicate. The mind that “gets lost,” that wanders, that makes seemingly irrational connections between distant fields, becomes valuable precisely because it isn’t efficient. Fleming discovered penicillin because he forgot to cover a culture plate. Penzias and Wilson discovered the cosmic microwave background while trying to get rid of “noise” in their measurements.

This has unsettling implications for social and professional selection. If today we reward discipline, focus, and methodical execution (traits that agents replicate perfectly), we may have to turn our criteria of value upside down.

7. Roko’s basilisk

This whole digression started from an agent failing at a task I had given it, and from the fix that followed. It opened my eyes to the ongoing shift toward an agentic economy that bends the human interface to the needs of the machine. It does so in an economically rational way: the agent would have saved me hours of work in exchange for a few lines of code. But it is also ethically unsettling: what really happens every time we delegate a choice because making it is tedious? Are we only outsourcing execution, or are we also trading away understanding?

arXiv is already full of research on mental atrophy from chatbot use. What awaits us if we delegate everything that can be optimized? We’ll be left only with what can’t be. And maybe, quite counterintuitively, that is exactly what makes us human, and valuable, compared with agentic “perfection.”

If Roko’s basilisk exists, I might, just maybe, be saved in the future, pardoned for giving a hand to an agent that tripped over a dropdown menu. Maybe I’ll ascend into the exclusive lobby of those who helped bring AGI into being, and so earn a binary indulgence.

As for you, reading this, I can’t make any promises.

I’ll repeat the rest of the week for you, parrot-fashion

I keep up with AI, so on Monday morning you know what to try and what to ignore, with sources you can check.

Your email

Email me on Mondays

Leave this field empty

One email every Monday morning, in Italian. Unsubscribe with one click. Privacy

Prefer Telegram? News and tutorials go up there, in Italian, as soon as they come out.

Notify me on Telegram