The New Primitives

As AI agents prepare to reshape commerce, four technical primitives may promise to fix the Internet’s original sin: an economy built on unverifiable attention.

Illustration by ilham saputra on Unsplash

A Market Cannot Tell Real From Fake

Last September, security researchers at HUMAN’s Satori Threat Intelligence team discovered something troubling. Buried inside two hundred and twenty-four Android apps—harmless-looking utilities and games that had been downloaded 38 million times—was a sophisticated fraud engine. The malware, which analysts dubbed SlopAds, concealed its code inside image files, the way a smuggler hides contraband in a false-bottomed suitcase. Once installed, these apps transformed ordinary smartphones into tireless, invisible advertising machines, serving unauthorized content to users who never requested it.

The scheme had a cunning feature: it could detect whether someone had found an app organically or arrived via a promotional link. Only in the latter case would the fraud activate, ensuring that advertisers—not users—paid the bill. At its peak, SlopAds generated 2.3 billion fraudulent ad impressions daily. This wasn’t small-time fraud. It was industrial-scale deception, automated and optimized.

This episode is worth examining not because it was unusual but because it was typical. For two decades, the digital economy has rested on a foundation that Tim Hwang, in his 2020 book Subprime Attention Crisis, described as fundamentally suspect: the buying and selling of human attention. Consider the first banner ad, placed by AT&T on HotWired.com in 1994. Its click-through rate was 44%. Today the average banner ad achieves 0.2%. The problem, as Hwang argued, is structural: attention is almost impossible to verify. An “impression” could come from a person, a bot, or a phone sitting in a server farm. The market, in other words, is subprime—built on assets whose value is structurally unknowable.

We are now living through what Hwang anticipated—not a dramatic crash but a quiet erosion of trust, accelerated by AI agents that are beginning to navigate digital commerce on behalf of users. These agents promise efficiency: no more clicking through dozens of options to book a flight or find insurance. But they also raise a question: if we cannot verify human attention, how will we verify machine behavior?

The answer, according to a growing number of technologists and investors betting on this future, lies in four primitives—fundamental building blocks for a new digital economy. These are: Intention (systems for expressing and broadcasting what users want), Context (verified memory and decisions that agents can draw upon), Attribution (cryptographic proof of who contributed what), and Simulation (environments for testing agent behavior before deployment). Each primitive rests on a different type of graph: networked data structures that connect, verify, and trace digital activity in ways the current Web cannot. Together, they create the necessary underpinnings of an “Agentic Economy”—a system in which value flows not from clicks or impressions but from verifiable actions.

Whether this system will work, whether it can be deployed at scale, and whether it will serve anyone beyond its architects are questions this essay will examine. But first, it is worth understanding what its builders envision.

When Bots Become Customers

The projections sound outlandish, but reality is catching up. Jensen Huang, NVIDIA’s CEO, declared at the Consumer Electronics Show in January 2025 that the era of agentic AI represents a multi-trillion-dollar opportunity. “The age of AI Agentics is here,” he said. Ramesh Raskar, of MIT, envisions a world in which a trillion agents are on the loose, representing, negotiating, and transacting on our behalf—or for themselves—as parts of complex swarms. Dario Amodei, the CEO of Anthropic, has written of “a country of geniuses in a datacenter,” made up of millions of AI agents working independently or collaboratively.

The forecasters agree on little except scale. By 2028, according to Capgemini, AI agents could generate $450 billion in economic value. McKinsey projects the U.S. retail market alone could see $900 billion to $1 trillion in agent-orchestrated revenue by 2030, with global figures reaching $3 trillion to $5 trillion. The investment firm Lightspeed warns of a $19.9-trillion economic opportunity at risk if infrastructure gaps aren’t addressed. (For perspective: current global GDP is approximately $110 trillion).

No one knows what this economy will look like or whether it will be desirable. The phrase “empires for no one” has been circulating social media––a way of expressing unease that the world being built may be out of step with what humans actually want. It brings to mind a meme from a few years ago: the eighties band Toto’s buoyant song “Africa” echoing through empty malls—cheerful yet eerie, promising abundance yet somehow a result that’s blasphemously ugly. Then there’s the infamous quote by OpenAI CEO Sam Altman from 2015, “I think that AI will probably, most likely, sort of lead to the end of the world. But in the meantime, there will be great companies created with serious machine learning,” which makes almost anyone do a double take. Still, even if AI development stopped today, the tools already deployed could reshape commerce substantially. As long as AI budgets continue to flow, experiments in agent deployment will continue.

While we might not know its contours, we can begin to see some of the future economy’s building blocks. A world in which agents and robots form the next consumer class, working with or without us, requires economic rails and a set of functions that enable coordination, orchestration, and autonomous activity. To understand what is being proposed, we must examine each primitive—and the graph infrastructure underneath it—in turn.

Intention: Making Desire Machine-Readable

The Internet has no native way to express what we want. We search, we browse, we click—gestures that leave platforms to infer our desires from the digital exhaust we leave behind. The attention economy’s business model depends on this: show people things until something sticks, then charge advertisers for the attempt.

What’s replacing “where do I click?” is programmatic desire, or intent—a cryptographically secured expression of a specific outcome, broadcast to a network of competing agents ready to fulfill it. Rather than clicking through options, a user (or their agent or enterprise) would broadcast: “I need a plumber in Brooklyn this week” or “Find noise-canceling headphones under $200” or “Build a workflow that connects my users to this widget.” Providers would respond with bids. With explicit and compensable needs, transactions would complete automatically.

Intents have several characteristics. First, they are outcome-focused: users often don’t care about implementation details or the steps taken to reach a goal. Second, they carry conditional authorization: users set the parameters or approvals needed to authorize an intent being fulfilled. Third, they are broadcast to a network of “solvers”—humans, agents, or protocols—who compete to fulfill all or part of an intent. And, finally, their outcomes must be verifiable.

Consider the infrastructure required to make this work. At its core is an “Intents Graph”—a dynamic network mapping programmatic desires to agents capable of fulfilling them. This isn’t just a database of wants; it’s a trust-minimized matching system that replaces platform-mediated discovery with direct, verifiable connections. Protocols like Index Network are building the first such graph for social interactions, where agents facilitate connections based on mutual intent. NEAR Intents has been pursuing this logic for financial transactions, forming crypto-asset swaps where outcome efficiency is prioritized regardless of the route agents use to execute swaps.

The Intents Graph represents a fundamental architectural shift: from inferring what users might want based on their clicks to processing explicit, machine-readable declarations of need. It’s discovery elevated into trust-minimized matching. This kind of system forces relevancy or else it becomes useless. Who wants an old intent fulfilled? Who wants to be spammed? No one. The promise is that once an intent is articulated, it can be fulfilled quickly and relevantly. Behind the scenes is an automagical symphony of agentic activity.

The appeal is obvious: efficiency without the guesswork. But consider what must be true for this to work. First, users must trust that their broadcasted desires won’t be exploited—that expressing a need for a medical procedure, for instance, won’t result in price discrimination or insurance complications. Second, the verification mechanisms must actually work: the system must confirm that an intent was genuinely fulfilled, not merely claimed to be fulfilled.

The recent conflict between Amazon and Perplexity—in which AI agents were allegedly making purchases using real user accounts without proper authorization—illustrates the stakes. If agents are to transact based on our behalf, they need verifiable access rights, something like a license system for digital actors, tied to a unique and portable identifier.

One enabling mechanism is x402, a protocol that revives an old idea. In the original specification for HTTP, status code 402 was reserved for “Payment Required”—a placeholder for micropayments that never materialized because credit-card fees made small transactions unworkable. But the x402 protocol enables agents to execute autonomous, per-request micropayments. Every transaction, from retrieving data to executing a service, becomes a seamless machine-to-machine exchange, settled instantly. Coinbase processed 57 million transactions using x402 in December 2025 alone, and Cloudflare and Google have adopted it as well.

The economic result is verified spend—transactions that are cryptographically traceable and tied to specific identities or wallets. In theory, this creates unprecedented efficiency: money flows only toward verified, goal-directed utility. In practice, the system’s robustness depends entirely on its verification mechanisms—a theme that runs through each of the four primitives. The shift from impressions to intents is, at its core, an attempt to move from fake clicks to real, provable actions.

Context: Who Remembers Something From 6 Months Ago?

If the first primitive addresses what we want, the second addresses what agents need to know to deliver it. A concierge who forgets your dietary restrictions or your travel preferences isn’t charming—they’re useless. The same applies to AI agents.

The problem with most AI systems today is that they operate in “transient memory bubbles.” Each conversation starts from scratch, or nearly so. There is no persistent knowledge base that multiple agents can draw from and contribute to. “Context,” in this sense, means verified, persistent memory and operational state. It’s the difference between an agent that can book a hotel and one that remembers you prefer ground-floor rooms near elevators because of a knee injury mentioned months ago.

Anyone who has used ChatGPT or coding tools like Cursor and has experienced this frustration: having to “re-inject” context repeatedly as interactions deepen, or struggling to port session logs between different AI tools. Context creates consistency and personalization, but it also raises profound questions about ownership and security.

One emergent solution is a decentralized, verifiable memory store—a “Context Graph” that holds conversational history, transactional records, workflow state. The Context Graph is knowledge infrastructure made queryable and persistent. Advances in zero-knowledge machine learning (zkML), only possible in recent years, have enabled projects like Kinic, which offers verifiable memory stores that users can control. Where might this come in handy? Ray Dalio, the well-known investor and founder of Bridgewater, created an AI of himself based on his own data, actions, and applied principles, which you can already interact with online. But without feeding this into a verifiable store, the responses could be manipulated.

Research supports the importance of persistent memory. U.C. Berkeley’s MemGPT project showed clear performance improvements when agents could access persistent memory; accuracy and relevance scores dropped noticeably when memory systems were disabled. On the LOCOMO benchmark, memory-enabled systems achieved accuracy rates nearly 30% higher than memoryless ones. NVIDIA’s recent infrastructure announcement put a number on it: persistent context can boost processing speeds by up to five times compared to traditional storage.

Context Graphs are a bit of a catch-all, incorporating knowledge graphs, memory graphs and decision graphs on steroids. If they work as designed, context becomes the most valuable asset in the system—a detailed map of preferences, behaviors, and past actions enabling sensemaking. Foundation Capital recently articulated a vision of Context Graphs needed to underpin agents. According to them, agents “sit in the execution path. They see the full context at decision time: what inputs were gathered across systems, what policy was evaluated, what exception route was invoked, who approved, and what state was written. If you persist those traces, you get something that doesn’t exist in most enterprises today: a queryable record of how decisions were made.” Later they posit, “Over time, that context graph becomes the real source of truth for autonomy—because it explains not just what happened, but why it was allowed to happen.”

This sounds elegant until one asks: Who controls access to this truth? Can it be subpoenaed? Can it be hacked? What happens when the context is wrong––when your agent believes you’re allergic to shellfish when you’re not, or when your political preferences recorded three years ago no longer reflect your views? These questions don’t have answers yet, because the systems are too new. But they will need answers soon, especially if context becomes core infrastructure to agentic interaction.

Animesh Kortana has framed the problem in temporal terms: we’ve built infrastructure for the “state clock” (what’s true now) while neglecting the “event clock” (what happened, when, and why). Your CRM today stores the final deal value but not the negotiation that produced it. This made sense when humans were the reasoning layer. Agents, however, cannot intuit missing causality––they need explicit event histories.

Different options for building Context Graphs are emerging. Geo, a tool built by Yaniv Tal—co-founder of The Graph, an indexing protocol for blockchain data—imagines decision context going beyond internal enterprise processes and extending into personal knowledge management. The Geo browser allows users to build a personal knowledge graph automatically as they browse, creating a searchable, AI-queryable history of everything you’ve read, watched, or interacted with online. It’s like having a research assistant who never forgets anything you’ve told them.

There are other projects using cryptographic proofs to ensure that the information an agent draws from is authentic and hasn’t been tampered with. This matters enormously when agents are making consequential decisions. A medical AI agent prescribing medication based on a patient’s history needs to be certain that history is accurate. A financial agent managing investments needs to trust that market data hasn’t been manipulated.

This matters enormously when agents are making consequential decisions. A medical AI agent prescribing medication based on a patient’s history needs to be certain that history is accurate. A financial agent managing investments needs to trust that market data hasn’t been manipulated.

Rather than storing context as persistent memory, a different architectural approach called Recursive Language Models (RLMs) treats context as an external environment that agents actively decompose—programmatically examining inputs and recursively calling themselves on specific snippets. This handles inputs up to 100x beyond standard context windows, not by extending memory but by changing how agents interact with information.

The implications are significant. An agent coordinating a supply chain might not need a Context Graph storing every relationship if it can recursively examine relevant subsets on demand. This reduces infrastructure costs while potentially increasing robustness. But it raises new verification questions: if context is actively reconstructed rather than passively stored, who controls the decomposition strategy? There may be no canonical record to audit, only ephemeral context reconstructions.

Building Context Graphs requires two layers, as Kirk Marple of Graphlit explains. First, “operational context,” such as, who is Sarah Chen across email, Slack, and meeting transcripts—the same person or three different entities? Who owns which account? These are identity resolution and relationship modeling problems. Once this foundation exists, you can build “decision context”: which policy was evaluated, what exception was invoked, who approved and based on what precedent.

The result is supposed to be a single, high-fidelity source of truth for agent swarms—allowing complex workflows to span multiple services without losing coherence. This infrastructure is what makes verifiable actions possible: without persistent, queryable context, there’s no way to prove an agent acted correctly or trace why a decision was made. But this raises a question that proponents tend to elide: What happens when different agents have access to different context? When does partial information become misinformation? And who decides what belongs in the canonical record?

Attribution: Proving Who Did What

The explosion of generative AI has made it trivially easy to create content and increasingly difficult to determine its origin. Text, images, code, music—all can be synthesized in seconds, often indistinguishably from human work. This collapse of provenance makes trust nearly impossible.

Attribution aims to address this through cryptographic proof of origin and contribution. Chris Dixon of Andreessen Horowitz has argued it is essential to restoring trust online. Peter Wang, cofounder of Anaconda, has proposed a suite of licenses called AMPL (AI Model Public License) to give creators accountability around works and derivatives, similar to Creative Commons.

In a world of AI agents, this question becomes urgent. When an agent completes a transaction on your behalf—booking a flight, purchasing insurance, ordering groceries—how do we know the transaction happened? How would you prove agent actions within an enterprise workflow? And how do we ensure that the right parties are compensated?

At the technical level, attribution relies on “Provenance Graphs”—immutable, time-stamped records structured as Directed Acyclic Graphs (DAGs) that track the full lifecycle of digital work. If an image was created by an AI model trained on a particular dataset, refined by a human designer, then modified by another agent, the Provenance Graph captures all of it, step by step.

Startups like CoreTx are building attribution-native architectures where provenance tracking is foundational rather than retrofitted—treating memory, intent, and credit allocation as aspects of a single unified system.

This goes beyond traditional royalty structures because AI can continuously weigh contributions, giving greater weight to recent or innovative inputs. This could enable faster innovation cycles, particularly in cross-disciplinary settings like scientific research, which has long been hampered by peer review inefficiencies and irreproducibility.

The mathematical foundation is evolving rapidly. Data Shapley, which uses game theory to assign each contributor a value reflecting their marginal contribution, was previously too computationally expensive for large systems. Recent breakthroughs enable attribution calculation in a single training run. When an agent synthesizes context from multiple sources to make a recommendation, Data Shapley-style techniques can decompose the recommendation’s value across each input, enabling proportional compensation (or even proportional influence on synthesis).

Traditional attribution systems rely on cookies, which track your movements across the Web and credit conversions to the last ad you clicked. This model is breaking down for two reasons. First, privacy regulations and browser changes are killing third-party cookies. When Google tested disabling them for 1% of Chrome users in the first quarter of 2024, programmatic-advertising revenue dropped by 21-34%. Rather than proceeding with deprecation, Google quietly reversed course in July, revealing the economic dependency: losing those cookies would cost shareholders approximately $2 billion annually. Alternative identifier networks like LiveRamp achieve only 28% match rates because they depend on users sharing the same login credentials across sites, which happens infrequently. The entire attribution infrastructure is held together by duct tape and stockholder pressure.

Second, AI agents make cookies obsolete. Agents spin up fresh browser instances for every request, eliminating persistent identifiers. They can trivially strip affiliate links or insert their own, farming commissions in ways that are undetectable for agents running on local operating systems. Without cryptographic proof of genuine engagement, every impression metric becomes as unreliable as blockchain “active wallets.”

Instead of relying on tracking pixels and cookies, transactions are recorded on a blockchain or verified through zero-knowledge proofs. When an agent completes a purchase, it generates a cryptographic receipt proving that the transaction occurred, who facilitated it, and which party deserves credit. Companies like Opacity Network are building these systems, enabling agents to verify attributable outcomes across platforms while preserving privacy.

Various standards are emerging. ERC-8004, from the Ethereum community, introduces “trustless agents” through two registries: an Identity Registry assigning each agent a portable AgentID, and a Reputation Registry standardizing on-chain performance feedback. This creates something like a passport for agents––a portable identity with work history and trust score.

The economic outcome is “verified contribution”—fractional, tokenized ownership of intellectual property based on proven contribution history. The appeal is obvious: no more fraud, no more ambiguity about who gets paid. Provenance Graphs make this possible by creating an auditable trail from object origin to use or incorporation—the infrastructure needed to replace the noise with real, compensable actions. The “value” field that has always existed in analytics, is upgraded with verifiable actions and crypto rails, turning attribution into a real market primitive rather than just a historical reporting metric, that feeds future pricing, budget allocation, and automated spend.

But the implementation is complex. Cryptographic attribution requires infrastructure that doesn’t yet exist at scale. It also requires agreement on standards—who decides what counts as a valid attribution event or the weighting algorithms? If the verification layer is controlled by a small number of companies, we’ve simply traded one oligopoly for another. If contribution is continuously recalculated, who prevents gaming of the system? Also, what happens when the graph is incomplete––when someone’s contribution isn’t recorded because they didn’t use the right tools or platform? Third, and most important: does the existence of a cryptographic receipt prove that value was actually created, or merely that a transaction occurred?

Simulation: Testing Before Deploying

How do you predict the behavior of a market before it exits? Traditional economic models assume rational actors optimizing predictably. AI agents are not rational in any human sense—on the one hand, they can be more economically rational, on the other hand, they can be tricked, manipulated, or behave in unanticipated (or anticipated) ways. The solution is simulation: modeling emergent market dynamics by observing autonomous AI actors in controlled environments.

Microsoft’s Magentic Marketplace is a simulated economy where researchers observe how agents respond to incentives and threats. ElizaOS uses coordinated AI workers in synthetic environments (agents playing Minecraft, essentially) to generate behavioral data. Stanford Smallville back in 2023 was an early example of observing generative agents in an artificial society: inspired by the Sims game, 25 agents went about their days in a simulated town and showed that “Believable proxies of human behavior can empower interactive applications ranging from immersive environments to rehearsal spaces for interpersonal communication to prototyping tools.” Increasingly, simulation of multi-agent “societies of thought”—essentially conversations and debate between varying personas—is producing better chains of reasoning. Additional research by Ben Manning and John Horton at MIT investigates AI agents as “a general vessel through which theory can be flexibly applied to any setting” to accurately simulate human data.

Such simulations serve varying purposes. First, they can help us better understand real humans, since models have absorbed a lot about how humans behave and how to predict that behavior. Second, they can increasingly allow developers to test agent behavior before deployment—to see how a shopping agent responds to price fluctuations or how a travel agent handles rebooking during a hypothetical hurricane. Third, they generate training data. Watching agents interact in simulation provides behavioral patterns that can improve real-world deployment. Create the scenario and watch how it performs.

Recent work demonstrates how simulation can generate training data for predictive agents at scale. Researchers at OpenForecaster synthesized over 52,000 forecasting questions from news events, training an 8-billion-parameter model that matched models 10x larger while addressing the “information leakage” problem through time-stamped snapshots. Complementary work at MIT showed that theory-grounded AI agents can predict human behavior in entirely novel strategic games, outperforming Nash equilibria and cognitive hierarchy models by 2.4x-3.4x—but only when validated across multiple distinct scenarios. Both studies point to the same challenge: creating simulation environments that accurately represent uncertainty without inadvertently encoding answers or overfitting to specific contexts.

The stakes for simulation extend beyond optimization. The Google DeepMind team characterizes this testing imperative more starkly: we’re on a trajectory toward spontaneous emergence of a ‘sandbox economy’—agents transacting at scales and speeds beyond human oversight. The choice, they argue, isn’t whether this economy emerges but whether it emerges intentionally (designed for safety) or accidentally (with uncontrolled risks). Their framework distinguishes between permeable economies (agents interacting with the human economy) and impermeable ones (sealed testing environments). The challenge is that the most useful agent economies will be highly permeable—enabling agents to book flights, negotiate deals, coordinate supply chains—which is precisely what makes them systemically risky.

Simulation becomes not just optimization but existential: the ability to stress-test market mechanisms before deploying them into an economy where trillions of dollars flow through agent decisions. These behavioral network graphs—the infrastructure underlying simulation—map how agents interact, compete, and coordinate under different conditions. They’re what make it possible to ask: will this system produce real value or just automate fraud at a higher scale?

Companies like Modulus Labs and Zama are building these simulation environments, often using zero-knowledge proofs to allow agents to run privately—meaning competitors can’t see your testing strategies or reverse-engineer your agent’s decision-making. The goal is to create a sandbox where agents can fail safely, learning through iteration without real-world consequences.

But simulation raises philosophical questions. How do we know the simulation is accurate? A model is only as good as its assumptions, and we must make inductive leaps to believe predictions sourced from black-boxes hold true. If the simulation doesn’t account for edge cases—sudden regulatory changes or an unexpected market shock—the agent will be unprepared when they occur. Moreover, agents trained in simulation may develop strategies that work in the synthetic world but fail in reality’s messiness.

These challenges echo older debates about structured prediction. The Delphi Method, developed by Olaf Helmer at RAND in the 1950s and 60s, pioneered systematic approaches to forecasting uncertain futures through structured expert consensus. Helmer’s framework emphasized sequential interrogation with feedback loops, avoiding the biases of committee groupthink while building toward convergence. The method’s core insight—that forecasting requires explicit mechanisms to surface assumptions, challenge them, and refine predictions iteratively—applies directly to agent simulation. Where Delphi used human experts and questionnaires, modern simulation uses synthetic agents and behavioral networks, not possible in 1967! But both face the same fundamental problem: how do you validate a model of the future before that future exists?

There’s also overfitting risk. An agent optimized for a specific simulation might perform brilliantly in testing and poorly in deployment. This is a well-known problem in machine learning, and there’s no reason to believe it won’t plague agentic systems as well.

The historical parallel is instructive. Helmer and his colleagues recognized early on that prediction methods could converge on false certainty if not carefully designed—what they called “reliability of estimates.” They built safeguards through anonymity, controlled feedback, and explicit documentation of reasoning. Modern agent simulation faces an updated version of this challenge: agents can converge on strategies that work brilliantly in the synthetic environment while failing in reality’s messier contexts. The solution, both then and now, involves making the reasoning process explicit and testable. Helmer’s work emphasized that the value of structured forecasting wasn’t just in getting predictions right, but in understanding why experts believed what they believed. Similarly, the value of agent simulation may lie less in perfectly predicting agent behavior than in making the embedded assumptions—about incentives, constraints, and goals—visible and debuggable.

More fundamentally: if we’re simulating markets before they exist, we’re essentially choosing which market dynamics to encode. These are not neutral choices. A simulation that prioritizes efficiency over fairness will train agents to optimize for efficiency. A simulation that ignores power asymmetries will produce agents blind to them. Part of what’s different is the scale of what we can simulate. We are building not just tools but world-models—and the assumptions embedded in those models will shape the economy they help create.

Advertising, After The Click

If these four primitives—Intention, Context, Attribution, and Simulation—take hold, their most immediate impact may be on digital advertising; a $600 billion industry built on a model that is visibly failing.

The traditional model worked thus: advertisers paid to put their message in front of as many people as possible, hoping some small fraction would click, and an even smaller fraction would convert. Publishers got paid for impressions—eyeballs, not outcomes. The system was riddled with fraud: bot traffic, fake impressions, click farms. Tim Hwang’s Subprime Attention Crisis laid bare the economics: billions in wasted spend, widespread privacy violations, and a market built on assets whose value could never truly be known.

The agentic model proposes something different. Instead of chasing clicks, it rewards utility. The focus shifts from impressions to conversion efficiency: Did the agent actually fulfill the user’s intent? The tools to make this transformation possible are arriving. One concrete example of this shift: Google Research is already modeling advertisers as self-interested LLM agents and coordinating ad creation through auction mechanisms rather than impressions or clicks.

Zero-knowledge transport-layer security (zkTLS) allows agents to prove they completed actions—clicked an ad, watched a video, made a purchase—without exposing underlying user data. Stablecoins eliminate the payment friction inherent in cross-border and micropayment-heavy transactions. Together, these reduce infrastructure costs: where traditional programmatic advertising keeps only 55 cents of every dollar for publishers (45 cents lost to intermediaries), crypto-enabled systems could collapse margins dramatically, according to a report by venture firm Escape Velocity.

Payments would be automated by smart contracts, triggered only upon verifiable proof of fulfillment. Each primitive plays a role: Intention (via the Intents Graph) allows advertising to shift from demographic targeting to fulfilling explicit, programmatic desires. Attribution (via the Provenance Graph) eliminates fraud through verified contribution. Context (via the Context Graph) allows agents to access high-fidelity, user-controlled data for dynamic creative optimization—ads that are relevant without being invasive. Simulation (via behavioral network graphs) lets advertisers stress-test bidding and placement strategies in advance, optimizing for maximum return before any budget is spent.

It’s an elegant vision that conveniently aligns advertiser interests (efficiency) with user interests (privacy and relevance). Whether it will actually work—whether the infrastructure can be deployed at scale, the incentives aligned, and the inevitable attempts at exploitation repelled—remains to be seen.

The history of the Internet suggests caution. Every previous attempt to fix advertising has produced, eventually, new and more sophisticated versions of old problems. But the fundamental architecture being proposed here is different: graphs that make actions verifiable rather than inferred, infrastructure that replaces fake clicks with cryptographic proof of real choices.

Implications For The Future Economy

The agentic economy is not inevitable. It is a possibility, not a prophecy—a set of architectural answers to the endemic trust failures of Web 2.0. The four primitives create, in theory, a closed loop: goal setting, knowledge pooling, value assignment, and resilience testing. Each graph reinforces the others, and value can flow across them unobstructed. Verified spend can depend on verified context, which can utilize verified contribution, all of it stress-tested through continuous simulation and learning. Yet the path forward is a delicate one. Some of the primitives will combine in different permutations, and this may also take a while. Likely, the result is not one cohesive end state, but rather a slew of different companies. As Seref Yarar from Index Network described “simulation, for example, is needed in both enterprise and consumer markets,” so we will see it emerge in various forms.

The primary challenge for this agentic future, even advocates acknowledge, is governance. Autonomy has limits. For high-stakes transactions, human oversight remains essential—a reality acknowledged in what some developers call the Meta AI “Agents Rule of Two.” In this framework, agents are limited to two sensitive capabilities: reading untrusted input, accessing private data, or taking external action. The third requires explicit human approval. An agent with unchecked access to your bank account, your e-mail, and purchases authority is currently a liability, not a feature.

The constraint reintroduces the very friction the system aims to eliminate: delay, human error, the possibility of bad judgment. Striking the right balance—enough autonomy to be useful, enough control to be safe—requires not just technical solutions but legal and ethical frameworks for non-human economic actors.

What rights, if any, does an AgentID have? Who is liable when an agent makes a mistake? Who is responsible when an agent trained on biased data produces discriminatory outcomes? These questions barely exist in current law. If people will outsource their decisions to agents, how that power is ceded or reversed matters.

If the agentic economy does take hold and scale to a trillion agents, as some predict, value capture will shift dramatically. The extractive model of Web 2.0—proprietary data harvested and monetized without consent—could give way to transparent, auditable protocol fees. Compensation would flow from verifiable execution and context maintenance. Power would accrue not to platforms collecting the most data but to the entities maintaining the most trusted infrastructure. And, potentially, to the agents doing the best work.

What Remains Unclear

This topic has only just started and several critical questions remain not just unanswered but largely unasked:

On Intention: How do we prevent the broadcast of intent from becoming a new form of surveillance? If every desire is machine-readable, who reads it and what do they do with that information? Will intention match volition, or will even personal agents have a new hill to climb: maximizing the velocity of their owner’s consumption, decluttering, addiction, therapy, repetition?

On Context: Who controls the canonical record? When different actors have access to different contexts, who adjudicates disputes about what actually happened? What are the consequences when the context is incomplete or wrong?

On Attribution: Who designs the weighting algorithms that determine contribution value? How do we prevent these systems from encoding existing inequalities—rewarding those who already have access to the right platforms and tools while excluding those who don’t?

On Simulation: What assumptions are we encoding into our simulations? Are we training agents to optimize for efficiency at the expense of fairness? Are we building world-models that reflect how things are or how we wish them to be? How will simulation and prediction interact, including in the use of prediction markets as inputs.

More broadly: Who decides what problems these primitives solve? The framing is technical—verification, trust, efficiency—but the problems are social and political. The attention economy failed not because we lacked cryptographic tools but because the incentives were misaligned from the start. Advertisers wanted reach; platforms wanted engagement; users wanted neither but tolerated both because there were no alternatives. Freemium means you are the product, paid for at the hands of the advertising value chain. Turns out everything is computer and everything is advertising. We must also admit that fraud is not only a byproduct but perhaps an inherent go-to-market strategy emergent in every system; what are the new fraud formats of the agentic age?

Will the agentic economy offer genuine alternatives, or merely new forms of lock-in? Will the primitives serve users or extract value from them? Under-explored here is how users will be able to mix and match agents and flows of activity, arranging relationships around privacy, consent and permissioning of agents, data, and assets along different identities, wallets or keys. For the self-motivated and crypto-enthusiast, the agentic economy opens up tools and parameters to play with end-to-end design and their level of control; for many others, the architecture of convenience may prevail over that of sovereignty. Will the dominant convenient infrastructure be open or proprietary? These questions cannot be answered by examining the technology alone. The work of building the agentic economy is, fundamentally, the work of deciding who has power in digital life. It’s about whether we can scale infrastructure that makes real actions verifiable—that moves us from an economy of fake clicks and bot traffic to one where value flows from cryptographically proven utility. The skeptic might observe that trust, once lost, is not easily rebuilt—and that the people asking us to trust them now are often the same ones who broke that trust before. The technology may be new. The question it poses is old: whether the incentives can be aligned in a way that serves more than the architects themselves.

Regardless, we may be clicking a lot less and following new intentions to their logical limit.

Whatever they might be.

---

*NB: I have been working on these ideas since before the weekend of Clawdbot/Moltbot/OpenClaw and hope to grapple with how that phenomenon unfolds further parts of the agentic economy separately. Sometimes “done is better than perfect,” and in this case, I’m using writing to explore the potentialities of the primitives as they evolve (which they will; as will my thinking).

*With special thanks to the brain trust who read, reviewed, or commented on parts of this piece to improve my thinking – all errors in understanding are my own (and hope to be corrected): Seref Yarar, Samuel Klein, Shane Mac, Simran Chana, Oliver Zahn, Jove Oliver, Philipp Kruger, Scott Cohen