The next contested layer of the digital economy will not be a new social network, app store or cloud platform. It will be the market infrastructure through which AI agents discover services, evaluate counterparties, agree terms, move money and trigger real-world obligations. That shift matters because the unit of economic activity is starting to change. In the consumer internet, software helped people search and click. In the agent economy, software will increasingly decide, procure and execute. The commercial consequences are larger than they first appear.
Much of today’s discussion about AI still assumes a familiar pattern: a human asks, a model answers. Yet the more consequential development is the rise of systems that can act with delegated authority. An agent can compare suppliers, choose a route for delivery, rebalance a marketing budget, renew software subscriptions, buy compute, hire micro-services, file routine paperwork or settle small claims. None of this requires science fiction. It requires only a chain of capabilities that already exist in rough form: machine-readable identity, access to payment rails, contractual wrappers, policy constraints, auditable logs and the ability to use tools across networks.
The difficult question is not whether agents can transact, but under what authority and with whose liability.
This is why the category of agent economy and markets deserves attention in its own right. It is not merely about smarter automation. It is about the rules of exchange when both sides of a transaction may be software, when pricing may be set dynamically by models, and when contractual commitments may be made at machine speed. The economics are promising. Search costs could fall, idle capacity could be matched more efficiently, and smaller firms could gain access to sophisticated procurement and treasury functions once reserved for large enterprises. But the governance problems are equally serious. If an agent spends badly, discriminates unlawfully, accepts predatory terms or exposes sensitive data, responsibility does not disappear into the model.
From assistants to economic actors
Software has long executed narrow commercial tasks. Algorithms trade securities, ad systems bid for impressions and industrial software orders replacement parts. What is changing is the breadth of delegation. Large language models, tool use and improved orchestration are allowing software to operate across previously separate systems: inboxes, procurement platforms, banking interfaces, customer records, logistics tools and legal templates. The result is not full autonomy in the philosophical sense. It is economic agency in the practical sense: the power to take actions that alter rights, obligations and cash flows.
This distinction matters. A chatbot that drafts an email creates little institutional risk. An agent that can accept a software licence, choose a payment method and trigger a recurring charge sits in a different category. Once software crosses from advice into commitment, the market must ask older questions in a new setting: who is the principal, what powers have been delegated, how are limits enforced, what evidence proves assent, and where can a harmed party seek recourse.
Why machine-to-machine commerce is plausible now
Three conditions are converging. First, more services are becoming programmable and machine-payable. APIs, usage-based billing and cloud-native operations have already normalised software buying software in constrained environments. Secondly, digital identity frameworks are maturing, even if unevenly across jurisdictions. Verifiable credentials, enterprise authentication standards and regulated trust services offer a path towards proving that an agent is acting for a recognised person or organisation. Thirdly, payment and settlement technologies are becoming more composable. Traditional payment rails, instant payment schemes and regulated forms of tokenised value all make smaller, more frequent and more conditional transactions easier to execute.
None of these layers is sufficient on its own. A market in agents needs interoperability between identity, policy and money. It also needs a legal wrapper. Electronic signatures, electronic trade documents and records retention rules become newly important when acceptance and execution happen without a human hovering over every step. The legal system has already adapted to digital documents in many areas. The unresolved issue is how far institutions are willing to extend delegated authority to software before requiring additional safeguards.
The promise: lower friction, broader participation
The difficult question is not whether agents can transact, but under what authority and with whose liability.
The attraction of the agent economy is straightforward. Firms waste enormous effort on search, comparison, routine negotiation, compliance checks and administrative processing. For a large company this is a cost centre; for a small firm it can be an existential tax. If a capable agent can compare vendors, parse service terms, flag anomalies and place low-risk orders within approved limits, a one-person enterprise can behave more like a mid-sized organisation. The same logic applies to households, charities, clinics, schools and municipalities with thin administrative capacity.
There is also a broader market effect. Many digital services are difficult to discover or too fiddly to procure in small increments. If agents can source exactly the amount of compute, translation, design, legal drafting or logistics support needed for a particular workflow, suppliers may be able to sell in smaller units and buyers may be able to assemble more tailored operational stacks. In theory, this should favour competition by reducing switching costs and making niche providers easier to find.
That optimistic case deserves to be taken seriously. Well-designed agentic markets could improve productivity without requiring every participant to become a specialist in procurement, security or financial operations. They could widen economic capability. For smaller actors, sovereignty does not mean doing everything manually. It means delegating safely without becoming dependent on opaque intermediaries that control identity, distribution and data.
The new bottlenecks: identity, authority and provenance
The central infrastructure problem is not intelligence but trust. A market of autonomous buyers and sellers is only as trustworthy as its systems for identity, provenance and recourse. A human can be challenged, called, sued or sanctioned. Software cannot, except through the entities that deploy and benefit from it. That means every meaningful transaction requires a chain of attribution: this agent represents this principal, under these permissions, for this period, with these spending limits, using this source of funds, subject to these audit requirements.
Identity for agents is not identity for persons
Machine identity should not be confused with personhood. The practical need is to establish verifiable linkage, not metaphysical status. An agent may require credentials proving that it is authorised by a company’s procurement policy, or that it is licensed to access regulated data, or that it can sign within a defined threshold. In many sectors, that will need to be granular and revocable. One agent may be permitted to buy cloud storage but not legal services; another may read invoices but not alter payment instructions.
Provenance is equally important. If an agent recommends or selects a supplier, what data did it use, what prompts or policies shaped the decision, and were those inputs tampered with. In digital markets, fraud often enters through identity spoofing, invoice manipulation and redirection of funds. Agentic workflows multiply the attack surface unless every step is strongly authenticated and logged.
Efficiency without controllability is not progress for firms, households or states.
Price formation in markets that think
Pricing in the agent economy will not look exactly like conventional software licensing. Some services will remain subscription-based, but many interactions will shift towards usage, outcome and event-based charging. A machine buyer may care less about branding and more about latency, reliability, permissions, jurisdiction, indemnity terms and the expected cost of failure. That could make markets more rational. It could also make them harsher.
When algorithms negotiate with algorithms, price discrimination may become more pervasive and less visible. Buyers may be profiled by urgency, historical tolerance for outages, inferred budget or the estimated sophistication of their policy controls. Sellers may dynamically alter contract terms, service levels or dispute procedures for different categories of agent. Human users often accept such opacity because the sums are small or the process is too tedious to contest. But if agents are making thousands of small commitments, those asymmetries can compound quickly.
A market of autonomous buyers and sellers is only as trustworthy as its systems for identity, provenance and recourse.
There is a parallel concern about collusion and concentration. If a small number of model providers, marketplaces or payment intermediaries become the standard gateways through which agents discover and contract with each other, they may shape market outcomes long before any formal competition case emerges. Ranking, default settings, proprietary identity schemes and closed policy languages can quietly determine who gets distribution and on what terms. The history of digital markets suggests that convenience and lock-in often arrive together.
Settlement is not a footnote
Economic action is only partly about decision-making; it is also about settlement. If agents are to transact frequently and in small amounts, payment systems must support programmability without forfeiting compliance, consumer protection or financial stability. In practice, this points to a hybrid world rather than a single rail. Conventional bank payments, card systems, real-time transfers and regulated tokenised instruments are all likely to coexist, each suited to different risk profiles and jurisdictions.
The key design question is how much discretion an agent should have over funds. A prudent architecture separates recommendation, authorisation and custody. An agent may be able to propose a purchase, trigger a low-value payment from a pre-funded budget, or request release from a treasury control layer. The more irreversible the payment, the stronger the need for explicit boundaries, authentication and reversible dispute processes. Markets built solely for speed will invite abuse. Markets built solely for caution will not scale. The institutional challenge is to find the acceptable middle.
Contracts written for humans will strain under machine execution
Commercial law is not absent from the agent economy. It is simply being tested in new ways. Existing doctrines of agency, authority, misrepresentation, negligence and unfair terms remain relevant. A company can delegate authority to an employee and can, in principle, delegate bounded authority to software. But many commercial contracts assume a human reader, a recognisable click-through process and a dispute over what a party knew or intended. When a model ingests dense terms and accepts them automatically, those assumptions become fragile.
One likely response is the growth of machine-readable commercial terms layered beneath conventional legal prose. Service levels, permitted uses, data-handling obligations, indemnities and termination triggers may need to be expressed in structured form so that agents can evaluate them before committing. That could improve clarity. It could also narrow the scope for productive ambiguity on which many commercial relationships depend. Law often works not because everything is perfectly specified, but because parties can interpret context and compromise. Machines are less good at that than their advocates suggest.
Liability will define adoption more than capability
For all the attention paid to model accuracy, the decisive factor for serious adoption may be liability allocation. If an agent books the wrong shipment, leaks confidential material, breaches sanctions rules or enters a contract outside its mandate, who bears the cost. Vendors will try to limit responsibility through disclaimers and capped warranties. Enterprises will insist on auditability and indemnity. Regulators will increasingly ask whether governance systems were proportionate to the foreseeable risk.
Here, the emerging AI governance frameworks are useful but incomplete. Standards bodies and regulators are right to stress risk management, human oversight, documentation and monitoring. Yet spending authority introduces a sharper edge. In many organisations, there is a large gap between allowing an employee to use generative AI for drafting and allowing an agent to commit budget or sign terms. The second requires not only model governance but financial controls, procurement rules and legal accountability. It is closer to delegated corporate power than to office productivity software.
Marketplaces will become regulators in all but name
Efficiency without controllability is not progress for firms, households or states.
Where agent-to-agent commerce scales, some form of organised venue is likely to emerge: directories, exchanges, broker layers or procurement hubs. These will not merely list services. They will set technical standards for identity, disclosure, dispute resolution, ranking, fee extraction and acceptable conduct. In effect, they will govern markets. The lesson from earlier platform eras is clear: rules embedded in infrastructure can matter more than formal law, especially in fast-moving cross-border sectors.
This creates a political economy problem. If marketplaces become the gatekeepers of trust, they can also become the gatekeepers of access. Small providers may find themselves complying with expensive verification schemes, accepting non-negotiable dispute rules or surrendering valuable data simply to be discoverable by machine buyers. Large firms may welcome this as a fraud-reduction measure. Smaller actors may see it as another layer of dependency. The tension is real. Security and openness do not line up neatly.
States will not be neutral observers
Governments have several reasons to care. Tax collection, consumer protection, sanctions enforcement, competition policy, digital identity and financial integrity all become harder when transactions multiply at machine speed across borders. Public authorities are therefore unlikely to permit a fully private constitutional order for agentic markets. They will push for traceability, accountability and standards interoperability, even if the methods differ between jurisdictions.
The divergence could be significant. The European approach is likely to emphasise documentation, risk classification and rights-based constraints. The United States may remain more sectoral and litigation-driven, with a stronger role for private ordering. Other jurisdictions may tie agent commerce more tightly to national digital identity systems or domestic payment infrastructures. For globally operating firms, this means the agent economy will not be one market but many overlapping ones. Compliance architecture will become a source of competitive advantage, but also a burden on smaller entrants.
The sovereignty question beneath the technology
At first glance, sovereignty may seem like an odd lens for markets built by software. In fact it is the central one. If individuals and small institutions are to benefit from agents, they must be able to delegate without surrendering control over identity, data, money and decision rights to a handful of dominant intermediaries. Otherwise the agent economy will merely reproduce the asymmetries of earlier platform eras, while adding the opacity of automated judgment.
Sovereignty does not mean isolation. Agents will depend on shared infrastructure, common standards and institutional trust. But it does mean portability of credentials, clear policy controls, inspectable logs, meaningful choice of payment and custody arrangements, and the ability to revoke or reassign authority without destroying one’s operating history. These are mundane design questions, yet they are where power accumulates.
What will matter by the end of the decade
By the late 2020s, the important divide may not be between companies that use AI and those that do not. It may be between organisations that can safely delegate commercial action to machines and those that cannot. The winners will not simply have better models. They will have better institutional design: cleaner authority structures, stronger identity management, clearer audit trails, more disciplined spending controls and more credible methods of redress when things go wrong.
The market narrative will be tempting to overstate. Not every purchase will be handled by autonomous agents, and many high-stakes deals will remain deliberately human. Trust, judgement and relationship still matter, especially where uncertainty is high. But routine commerce is vast, and it is routine commerce that often determines who can operate efficiently. In that domain, agentic systems are likely to become normal.
The foundational question for this category is therefore simple. As software starts to buy, sell and commit, what kind of market architecture will preserve accountability, contestability and institutional sovereignty rather than erode them. The answer will not come from model benchmarks alone. It will come from the slower work of law, payments, standards, governance and market design. That is where the real politics of the agent economy now sits.
