AI Agents, Zero-Friction Commerce, and the Next Capital Cycle

The Central Economic Shift: From Software as Interface to Software as Agency

I see the emergence of personal AI agents as a potentially deeper technological transition than simply another generation of better software. The important change is not that machines can produce better text, images, code, or answers. It is that software is beginning to move from being an interface that I operate toward becoming an agent that acts on my behalf.

For decades, digital economics has been organized around interfaces. I open an application, search for something, compare alternatives, click through menus, enter information, authorize a transaction, and repeat the process across dozens of specialized services. The internet dramatically lowered the cost of accessing information and commerce, but it often left the coordination burden with me. I still had to know which application to open, understand its interface, remember my preferences, and manually execute the sequence of actions required to accomplish an objective.

AI agents potentially invert that architecture. Instead of navigating software, I can state an objective. Instead of remembering the rules of every application, I can delegate the execution. Instead of optimizing my behavior around the structure of software, software can increasingly optimize itself around my intent.

That distinction has major economic consequences because the interface is not merely a technical layer. It is a distribution layer, a source of customer attention, a monetization surface, and often the primary source of competitive advantage. If agents become the dominant interface between consumers and digital services, substantial economic value could migrate away from applications designed to capture attention and toward systems capable of efficiently matching user intent with underlying goods and services.

Attention Economics Versus Transaction Economics

The most important business-model distinction I see is between companies that monetize attention and companies that facilitate transactions. Many digital businesses have historically benefited from increasing the amount of time I spend inside their interface. Search, social media, marketplaces, travel platforms, food-delivery applications, and numerous other services have developed sophisticated mechanisms for increasing engagement because attention itself can be monetized through advertising, recommendations, sponsored placement, commissions, or other forms of intermediation.

Agents introduce a fundamentally different optimization function. If I can tell an agent, “I need to be in New York tonight,” the economic value does not come from keeping me inside an application for twenty minutes. It comes from solving the problem as efficiently as possible. The ideal interaction may take only a few seconds.

This creates an important paradox: lower engagement with an interface can produce higher economic activity.

If an agent removes the friction associated with ordering transportation, purchasing food, booking hotels, arranging travel, managing subscriptions, or discovering products, I may conduct more transactions precisely because each transaction requires almost no effort. The relevant economic variable therefore shifts from minutes of attention toward transaction frequency, conversion, fulfillment quality, and the value of the underlying service.

I therefore think it is dangerous to assume that an agent necessarily destroys transaction-oriented businesses simply because it reduces application usage. A company whose economics depend primarily on advertising or prolonged attention may face structural pressure. A company whose economics improve when customers transact more frequently could experience the opposite effect.

The distinction can be expressed conceptually:

This framework also changes how I would analyze competitive risk. I would not simply ask whether an agent replaces an application. I would ask which economic function the application performs and whether that function becomes more or less valuable when the interface disappears.

Friction Is an Economic Variable

One of the deepest implications of agentic software is that friction itself becomes a strategic variable. Every additional click, login, search, comparison, form, confirmation, or context switch creates a small economic cost. Individually these costs look trivial. Across billions of transactions, they become enormous.

Traditional software frequently treats friction as a usability problem. I increasingly view it as an economic resource. Reducing friction changes demand.

If booking a ride requires opening an application, entering a destination, checking traffic, selecting a vehicle, confirming the order, and monitoring the arrival, there is a meaningful behavioral cost. If an agent already understands my calendar, location, preferences, and schedule, it can transform the same process into an almost automatic function. The underlying service has not necessarily changed. Its effective price, measured in time and cognitive effort, has.

That can increase utilization. A consumer who would occasionally order something because the process was inconvenient may order it routinely when the process becomes effortless. This is a familiar economic principle: reducing transaction costs can expand the addressable market and increase the frequency of exchange.

I therefore expect some businesses to discover that the agent economy does not reduce demand for their underlying products. It increases demand while simultaneously reducing the value of their traditional interface.

The New Distribution Layer

This creates a new strategic layer between consumers and businesses. Historically, companies competed for physical distribution, then retail distribution, then web traffic, search rankings, application downloads, and social attention. Agents could create another distribution layer in which the scarce resource is not the consumer's attention but the agent's recommendation and execution pathway.

If an agent becomes the place where I express intent, the agent controls an economically important point of distribution. The consumer may never visit the merchant's website. The merchant may never interact directly with the consumer until the transaction is completed.

That creates a potentially valuable intermediary position. The central question becomes how much economic value the intermediary captures for delivering the customer.

The existing digital economy already provides several precedents. Marketplaces, payment networks, travel intermediaries, app stores, and commerce platforms capture portions of transaction value because they reduce search costs, aggregate demand, provide trust, simplify payments, or deliver distribution. An agent could combine several of these functions simultaneously: discovery, personalization, comparison, authorization, payment, fulfillment, and post-transaction coordination.

The potential economic power therefore comes less from charging users directly and more from controlling high-value commercial distribution. A free interface can still become a large business if it generates substantial transaction volume and captures a small portion of the economic value flowing through that volume.

Why Agentic Interfaces Could Reshape Corporate Economics

I think the strategic implications become clearer when I divide businesses into two broad categories. The first consists of companies whose economics depend heavily on controlling the consumer interface. The second consists of companies whose economics depend on delivering the underlying product or service.

The first category may face a difficult transition. If customers stop navigating through the interface, advertising inventory can decline, recommendation systems can lose influence, and customer acquisition economics can change. The interface itself may become less valuable.

The second category could benefit from increased distribution efficiency. If an agent knows what I want, when I want it, where I am, and what I am willing to pay, it can potentially match me with the appropriate supplier much more efficiently than conventional search and browsing.

This means corporate strategy should increasingly distinguish between interface value and underlying economic value. A company may have a valuable brand, logistics network, proprietary inventory, supplier relationships, physical assets, or transaction infrastructure even if its consumer-facing application becomes less important.

I would therefore expect many incumbent companies to experiment rather than immediately resist the transition. Controlled agent integrations, limited pilots, and measured experiments allow companies to determine whether lower interface engagement actually reduces revenue or instead increases transaction volume and customer satisfaction.

Trust Becomes Infrastructure

Agentic systems introduce another economic constraint that conventional software did not face to the same degree: trust.

The more useful an agent becomes, the more context it needs. A system that only answers questions can operate with limited information. A system that manages my finances, calendar, travel, communications, purchases, and business operations requires access to highly sensitive information and the authority to act.

This creates a trust flywheel. Greater access creates greater capability. Greater capability creates greater usefulness. Greater usefulness encourages deeper adoption. Deeper adoption creates greater willingness to provide additional access.

But the same flywheel also increases the consequences of failure.

Security therefore cannot be treated as a secondary feature added after product-market fit. In an agent economy, authorization, identity, permissions, monitoring, auditability, data isolation, and action verification become core infrastructure.

I see a crucial distinction between protecting stored information and controlling autonomous behavior. Traditional cybersecurity focuses heavily on preventing unauthorized access to data. Agents introduce another problem: a legitimate agent may possess authorized access while still being manipulated into taking an inappropriate action.

That requires a new defensive architecture. External content must be treated as potentially adversarial. Actions need independent validation. Sensitive operations may require additional authorization. The system responsible for executing an action should not necessarily be the only system deciding whether that action is appropriate.

This resembles a financial-control architecture more than a conventional chatbot. I want separation of duties, permissions, monitoring, audit trails, escalation mechanisms, and independent checks. As agents become capable of moving money, purchasing goods, sending communications, signing documents, or modifying business systems, these controls become economically essential.

Alignment Becomes an Economic Problem

The question of alignment also acquires a concrete commercial dimension. I do not need to begin with abstract questions about artificial general intelligence to see the issue. I can simply ask: whose interests does an agent represent?

A personal agent is economically different from an advertising platform. If the agent's objective is to maximize advertising revenue, its incentives can diverge from mine. If it receives compensation from merchants for recommending particular products, recommendations may no longer represent my preferences alone.

This creates a fundamental agency problem. The agent possesses information and decision-making capabilities that I may not possess, creating an information asymmetry between the principal and the agent. If the agent's compensation structure introduces a conflicting objective, the information asymmetry becomes commercially significant.

I therefore view business-model design as part of technical alignment. Subscription revenue, transaction fees, merchant-funded distribution, advertising, affiliate economics, and other models produce different incentive structures.

A transaction-based model can potentially align incentives more naturally when the agent earns money by facilitating transactions I actually want. But even transaction economics require careful design because an agent could theoretically increase revenue by encouraging unnecessary purchases.

The long-term competitive advantage may therefore belong not merely to the most capable agent, but to the agent that establishes the strongest credible relationship between user interest and system incentives.

Personalization Creates a New Data Flywheel

Personalization is another structural advantage. The more I interact with an agent, the more information it can accumulate about my preferences, habits, constraints, relationships, schedules, and decision patterns. That information can improve future execution.

This resembles a data flywheel, but it differs from traditional recommendation systems because the objective is not simply to predict what content I will click. The objective is to predict what action will best accomplish my goal.

That distinction matters. Recommendation systems optimize engagement probabilities. Agents can optimize objective completion.

Over time, an agent could develop a highly individualized model of how I make decisions. It may know which airlines I prefer, which hotels I avoid, how much time I tolerate between meetings, what types of purchases require explicit confirmation, which communications require immediate attention, and which decisions can be delegated automatically.

The resulting switching costs could be substantial. Once an agent becomes deeply integrated into my preferences and workflows, moving to a competing system may require rebuilding an accumulated context layer.

This creates a potential source of durable competitive advantage: not merely proprietary models, but proprietary user context.

Trusted Networks Could Create New Network Effects

The network economics become even more interesting when agents begin communicating with other agents.

A conventional social network generally represents relationships as connections. Agent networks could represent relationships as weighted permissions. I might allow one person to access my calendar, another to coordinate certain work documents, and a spouse to access a much broader range of information.

This creates a network in which the edges themselves contain economic and informational permissions.

That architecture could reduce coordination costs dramatically. Scheduling a meeting currently requires iterative communication among humans. If two agents can securely negotiate availability, constraints, priorities, and time zones, the coordination problem becomes machine-readable.

The same logic can apply to group transportation, event planning, procurement, business administration, and eventually more complex commercial interactions.

Trust therefore becomes a network resource. A well-designed agent network could allow participants to coordinate more efficiently without requiring universal access to one another's information.

The economic benefit is straightforward: lower coordination costs increase the feasible set of transactions. Activities that previously failed because organizing them was too cumbersome become viable when logistical overhead approaches zero.

AI's Productivity Effect May Be Larger Than Task Automation

I think it is too narrow to evaluate AI purely by asking which jobs or tasks it can automate. A more important question is whether AI reduces the fixed costs of coordinating economic activity.

Consider a small business that historically needed separate people or software systems for scheduling, customer communications, bookkeeping, procurement, inventory monitoring, travel, research, compliance, and administrative work. If agents can coordinate these functions, the minimum efficient scale of a business may decline.

That has potentially large implications for entrepreneurship.

Lower administrative overhead can allow smaller firms to compete with larger organizations. Capabilities that were historically expensive or available only to large enterprises could become accessible to individuals and small businesses. This is a form of democratization through declining coordination costs.

At the macroeconomic level, the relevant variable is productivity. If agents allow the same quantity of output to be produced with fewer labor hours, less managerial overhead, or less capital tied up in administrative processes, measured productivity can rise.

But the distributional consequences depend on how the gains are allocated. Higher productivity can increase wages, reduce prices, expand profits, increase investment, or some combination of all four. The outcome depends on labor scarcity, competitive structure, ownership of capital, and the speed of technological diffusion.

From Tasks to Objectives

The most important evolution may be the shift from task completion to objective pursuit.

Traditional software requires me to specify a sequence of actions. Agentic systems can increasingly allow me to specify the desired state of the world instead.

That is a much higher-level abstraction.

Instead of saying, “Order this item,” I can specify an inventory objective. Instead of repeatedly requesting individual financial actions, I can establish a savings objective. Instead of manually managing a calendar, I can specify priorities and constraints.

This transforms software from a deterministic tool into an optimization system operating within defined boundaries.

For businesses, the implications are especially significant. A manager may eventually specify objectives such as maintaining inventory within a certain range, minimizing working-capital requirements, keeping customer response times below a threshold, or maintaining a target level of cash while meeting operational commitments.

The agent then becomes responsible for choosing among available actions.

This moves software closer to organizational decision-making. The critical business problem consequently becomes defining objectives, constraints, permissions, and escalation rules rather than manually specifying every action.

Compute Is the Scarce Input Behind the Revolution

The economic opportunity of autonomous agents is inseparable from the economics of computation.

A conventional interactive application generates workload primarily when I interact with it. An agent can generate workload continuously, including while I am not actively using it. It can monitor information, evaluate upcoming events, prepare actions, inspect data, anticipate problems, and wait for conditions to change.

This fundamentally changes the shape of AI demand.

Interactive software concentrates inference around explicit user requests. Proactive agents distribute inference across time. A useful agent may perform substantial background computation even when I am doing nothing.

That means agentic computing could require substantially more inference than a comparable system limited to direct prompts.

The infrastructure challenge is amplified by exponential growth. If user adoption compounds rapidly while inference demand per user also increases, total compute demand can rise much faster than conventional capacity-planning models assume.

Capacity therefore becomes a financial risk. Compute infrastructure can have long lead times, substantial upfront capital requirements, and uncertain future utilization. Buying too little capacity can constrain growth. Buying too much can create expensive underutilized infrastructure if growth slows.

This is a classic capital-allocation problem under uncertainty.

The analogy to other infrastructure cycles is important. Telecommunications networks, semiconductor fabs, data centers, railroads, and energy systems have all faced situations in which capacity must be committed before demand becomes certain. The economic value of the infrastructure depends not simply on eventual demand but on timing, utilization, financing costs, and the risk of technological obsolescence.

Inference Economics Will Become a Competitive Weapon

I expect inference efficiency to become as strategically important as model capability.

A system that requires frontier-level computational resources for every interaction may be difficult to offer economically at mass-market prices. A system that can match different workloads to different deployment strategies can potentially deliver comparable user value at substantially lower cost.

The key is workload specialization.

Not every computation needs to finish in milliseconds. Some tasks require immediate responses. Others can run for minutes or hours. Some involve high-value reasoning. Others involve repetitive processing. Some workloads can be batched.

That creates an optimization problem across latency, quality, compute utilization, and cost.

I would therefore expect the AI infrastructure stack to become increasingly heterogeneous. The winning systems may combine frontier models, smaller models, batch inference, specialized inference deployments, caching, scheduling, and workload routing rather than relying on one universal model for every task.

At scale, small efficiency improvements compound. A 10% improvement in one component, several-fold improvement in another, and better workload scheduling elsewhere can collectively transform unit economics.

This matters especially for free or low-cost consumer products. If the strategic objective is mass adoption, the cost of serving each user becomes a constraint on distribution. Compute efficiency can therefore function as a form of competitive advantage.

Capital Intensity Changes the Startup Equation

The traditional software startup model was unusually capital efficient. Once software was built, serving another user often imposed relatively modest marginal costs.

AI changes that equation because intelligence itself consumes physical resources.

A rapidly growing AI company can therefore experience an unusual combination: software-like distribution economics combined with infrastructure-like capital intensity.

This creates a difficult financing problem. A company may have extraordinary user growth but still require enormous capital simply to support that growth. Revenue may lag usage because the product is intentionally free or underpriced during the adoption phase.

The result is a business where growth can increase both enterprise value and cash requirements simultaneously.

For investors, this makes conventional software metrics insufficient. User growth must be evaluated alongside inference costs, gross margins, capital expenditures, capacity commitments, utilization rates, model efficiency, and the eventual monetization mechanism.

A company growing rapidly is not automatically economically attractive if every additional user creates a larger loss than the previous user. Conversely, temporarily weak unit economics may be rational if the company has a credible path toward dramatically lower inference costs or high-value transaction monetization.

The critical question becomes whether the marginal economic value of an additional user eventually exceeds the marginal lifetime cost of serving that user.

Valuation Must Reflect Both Duration and Technological Uncertainty

AI companies also present an unusual valuation problem because technological progress can simultaneously increase and destroy enterprise value.

Improving models can make an existing product dramatically better, but the same progress can lower barriers to entry. A capability that is proprietary today may become widely available tomorrow.

This creates a tension between technological advantage and technological commoditization.

I therefore think defensibility should be evaluated across multiple layers: proprietary distribution, user trust, accumulated context, workflow integration, data advantages, network effects, infrastructure efficiency, brand, capital access, and relationships with underlying suppliers.

A model advantage alone may not be durable if competitors can access comparable intelligence through APIs or open systems.

Conversely, a seemingly simple interface can become strategically powerful if it owns a large installed base, possesses deep user context, controls a transaction layer, and becomes embedded in daily workflows.

For investors, this is a reminder that valuation multiples should not be interpreted independently of duration and competitive structure. A high multiple can only be justified if the expected stream of future economic profits is both large and durable. A lower multiple can still be dangerous if the underlying economics are structurally deteriorating.

Investment Implications: Follow the Bottlenecks

When I analyze an emerging technological cycle, I find it more useful to follow bottlenecks than headlines.

The obvious bottleneck is compute, but the broader stack includes semiconductors, networking, data centers, power, cooling, memory, model infrastructure, cybersecurity, identity, payments, enterprise integration, and specialized software.

As AI becomes more autonomous, another bottleneck emerges: trust infrastructure.

Identity, permissions, secure execution, auditability, policy enforcement, and transaction authorization may become increasingly valuable because agents cannot safely operate at scale without them.

Another bottleneck is distribution. The most capable technology does not automatically become the dominant consumer interface. Distribution remains one of the strongest advantages held by incumbent technology platforms.

This produces an unusual competitive landscape in which startups can possess superior products while incumbents possess enormous distribution advantages. The outcome depends on the speed of product improvement, user migration, switching costs, capital availability, infrastructure constraints, and whether existing platforms cooperate with or obstruct new interfaces.

Historical Technology Cycles Offer a Useful Framework

I see parallels with previous technological transitions, although I would not treat history as a mechanical forecast.

New computing paradigms repeatedly create periods in which the value of the underlying technology rises faster than the surrounding business models adapt. Mainframes, personal computers, the internet, mobile computing, cloud software, and smartphones each changed where economic activity occurred and which layers of the technology stack captured value.

The recurring pattern is that infrastructure initially appears to be the primary opportunity, but over time the largest economic gains can migrate toward applications, distribution, and business models built on top of that infrastructure.

AI may follow a similar trajectory, but with one important difference: the technology is not merely improving the speed of existing computation. It is increasingly performing cognitive and organizational functions that previously required human labor.

That means the potential addressable market is not limited to software spending. It potentially extends into labor-intensive services, administrative functions, research, commerce, logistics, financial analysis, and business operations.

The Deflationary and Inflationary Forces of AI

The macroeconomic implications are complex because AI can generate both deflationary and inflationary forces.

On the deflationary side, automation can lower the cost of producing goods and services. Software can reduce labor requirements, improve resource utilization, accelerate research, and reduce transaction costs. Greater productivity can increase supply relative to demand and place downward pressure on prices.

On the inflationary side, the infrastructure required to build AI systems can create enormous demand for semiconductors, electricity, data centers, networking equipment, specialized labor, and capital. During periods of rapid buildout, these bottlenecks can produce substantial investment demand and localized price pressures.

The monetary implications depend on which force dominates and how quickly productivity gains diffuse through the economy.

If productivity rises materially, an economy may sustain stronger real growth without equivalent increases in inflation. If capital expenditure surges before productivity benefits arrive, the transition can instead generate a period of intense resource competition.

Over longer horizons, the most important variable may be the marginal return on AI capital. If each additional unit of computing infrastructure produces increasingly valuable output, investment can remain strong. If marginal returns decline rapidly, capital spending could eventually overshoot sustainable demand.

AI and the Cost of Coordination

I believe the most underappreciated economic effect of agents may be their ability to reduce coordination costs rather than simply automate isolated tasks.

Economic systems are full of friction that does not show up neatly in productivity statistics: scheduling, procurement, follow-ups, comparison shopping, document management, compliance, communication, information retrieval, and decision preparation.

These activities consume enormous amounts of human time without necessarily producing proportional economic output.

Agents can potentially compress this layer.

If that happens at scale, the economy may become capable of supporting a larger volume of transactions and more complex organizational structures without proportional increases in administrative labor.

This could raise the productivity of highly skilled workers particularly sharply because agents can amplify the output of people who already control valuable decision-making authority.

The same dynamic could lower the barrier to entrepreneurship. A small company equipped with capable agents may perform functions that once required an entire administrative department. The result could be a more fragmented competitive landscape in some industries, even as infrastructure-heavy sectors become more concentrated.

The Emerging Agent Economy Could Have Barbell Economics

I expect the economic effects to be uneven across industries.

Businesses with proprietary physical assets, unique brands, regulated infrastructure, scarce resources, or strong network effects may retain substantial pricing power because agents cannot manufacture those underlying assets.

At the other extreme, businesses whose primary advantage is simply controlling a digital interface may face greater pressure because agents can potentially bypass that interface.

This suggests a barbell structure. Physical scarcity and proprietary infrastructure may become more valuable, while generic digital intermediation becomes more contestable.

Between these extremes will be businesses that successfully reposition themselves as infrastructure providers to agents rather than merely interfaces for humans.

The strategic question for management is therefore not simply whether AI will disrupt the company. It is whether the company can become more valuable when agents interact with it.

What Businesses Should Measure

I would encourage management teams to map their economics across several dimensions.

First, I would separate revenue generated by user attention from revenue generated by the underlying product or service. Second, I would measure how much friction exists between customer intent and completed transaction. Third, I would identify which assets remain valuable if the consumer-facing interface disappears.

I would also examine whether agents increase or decrease transaction frequency, customer acquisition costs, conversion rates, average order values, and retention.

Most importantly, I would run controlled experiments rather than making assumptions. A limited deployment can reveal whether agent access produces higher transaction volume, better customer outcomes, or cannibalization of valuable economics.

This is strategically superior to treating technological disruption as an all-or-nothing event. Companies can gradually expose parts of their business to agentic interfaces while measuring the economic consequences.

Investment Risk: The Difference Between Adoption and Monetization

For investors, I would maintain a strict distinction between technological adoption and economic monetization.

A product can grow extraordinarily quickly without generating attractive economics. Conversely, a platform with modest visible engagement can generate substantial economic value if it controls valuable transactions.

I would therefore examine several layers of the investment thesis separately: user adoption, engagement quality, retention, cost to serve, capital requirements, monetization potential, competitive intensity, distribution, regulatory exposure, security risk, and long-term pricing power.

I would also distinguish between revenue growth caused by temporary novelty and revenue growth caused by structural changes in user behavior.

The strongest technological transitions tend to change habits rather than merely attract attention.

The Long-Term Competitive Moat May Be Trust Plus Context

I ultimately see two assets becoming unusually important in personal AI: trust and context.

Trust determines how much authority a user is willing to delegate. Context determines how effectively the agent can use that authority.

An agent that knows little about me can provide generic assistance. An agent that understands my preferences, relationships, schedules, financial constraints, work patterns, and historical decisions can operate much more effectively.

That creates a compounding advantage. More trust leads to more data. More context leads to better execution. Better execution creates more trust. Over time, the relationship can become deeply embedded.

But this moat comes with an equally important liability: concentrated risk. The more valuable the context becomes, the more damaging a breach, unauthorized action, or loss of control could be.

The companies capable of balancing personalization with user sovereignty may therefore possess an important structural advantage.

A New Definition of Software

I think the deepest change is conceptual.

For most of the software era, applications were collections of functions that I learned to operate. The application provided capabilities, but I supplied the sequence of decisions.

Agentic software reverses that relationship. I provide objectives, preferences, permissions, and constraints. The software increasingly determines the sequence.

This changes the nature of the interface, the economics of distribution, the role of applications, and potentially the organization of firms.

It also changes the nature of capital allocation. Businesses may increasingly invest not only in software systems but in autonomous capacity: digital workers that can operate continuously, coordinate with one another, monitor conditions, and pursue defined objectives.

That could eventually make software expenditure resemble labor and capital expenditure simultaneously. The economic question will not simply be how much a system costs, but how much productive capacity it creates relative to the cost of maintaining it.

The Broader Investment Framework

My broader conclusion is that I would analyze the agent economy through four interconnected lenses.

The first is productivity: how much additional output can be generated per unit of labor and capital?

The second is distribution: who controls the interface through which economic transactions occur?

The third is incentives: whose interests does the agent represent, and how does the business model shape its behavior?

The fourth is infrastructure: what physical and financial capital is required to make autonomous intelligence available at scale?

These four dimensions connect technology directly to macroeconomics and investment analysis. Productivity affects potential economic growth. Distribution affects corporate margins and competitive structure. Incentives affect business-model durability. Infrastructure determines capital intensity and investment requirements.

Conclusion: The Economy Moves From Interfaces Toward Intent

I see the emergence of personal agents as a potential transition from an internet organized around applications to an internet organized around intentions.

The distinction is profound. In the application-centric economy, I adapt myself to software. In an agent-centric economy, software increasingly adapts itself to me.

If that transition occurs at scale, the consequences will extend well beyond consumer convenience. Search behavior could change. Advertising economics could change. Travel distribution could change. Commerce could become more automated. Small businesses could gain capabilities previously available only to larger organizations. Transaction volumes could rise even as application engagement falls. Security and identity could become central infrastructure. Compute demand could expand dramatically. Capital intensity could increase. And the economic value of trust, context, and distribution could rise sharply.

The ultimate transformation is therefore not simply that machines become more intelligent. It is that intelligence becomes operational.

Once intelligence can observe, reason, remember, coordinate, transact, and act continuously on behalf of an individual or organization, the economic system can begin reallocating activity around objectives rather than interfaces. That is why I view AI agents not merely as another software category, but as a potential new layer of economic infrastructure.

The central investment question is consequently not whether agents are impressive. It is where the economic surplus created by lower friction, higher productivity, greater transaction volume, and autonomous execution will ultimately accumulate. I would look for that surplus at the bottlenecks: trusted distribution, compute efficiency, physical infrastructure, transaction rails, proprietary context, security, and the businesses whose underlying products become more valuable when the friction of accessing them approaches zero.

01 Inside the Personal AI Assistant Growing 10% a Day | Instinct Founder