The usual account of the agent economy begins with labour substitution. Software, on this view, drafts reports, triages support tickets, books travel and writes code. Useful though that frame is, it may prove too narrow. By mid-2026 the more consequential development is that agents are starting to occupy a different institutional role: not merely as instruments used by firms and households, but as operative counterparties inside markets. They request quotes, compare terms, trigger purchases, manage inventories, route workloads, negotiate service levels and, increasingly, determine when a dispute should be escalated to a human or a formal process.
This is an unexpected entry point because it shifts attention away from model capability and towards market design. Economies do not run on intelligence alone. They run on counterparties, contracts, payments, identity, trust, recourse and competition. Once software agents are delegated authority to perform commercially meaningful acts, old distinctions between user, tool and trader become less useful. The practical question is no longer whether an agent can compose a persuasive email. It is whether institutions built for human-paced exchange can cope when millions of machine actors search, bargain, accept and contest obligations continuously.
That question has broad implications. It affects procurement, logistics, cloud computing, financial services, advertising, energy management and professional work. It also affects law and regulation in a less obvious way. The principal challenge is not to decide whether an agent is a legal person; in most jurisdictions it is not. The challenge is to allocate responsibility when legally significant decisions are made by software operating under delegated authority, often across several firms, jurisdictions and layers of infrastructure. In that world, the bottleneck is less reasoning quality than commercial legibility.
From tools to market actors
Software has acted on behalf of people for decades. Algorithmic trading, dynamic pricing and ad auctions are hardly new. What is different now is the widening of machine agency into ordinary commercial workflows. A procurement system that once required static rules can now ingest a policy, survey suppliers, rank trade-offs, negotiate narrow parameters, redline standard clauses and prepare a recommendation in one loop. A consumer-facing assistant can not only search, but hold preferences, compare subscription terms, co-ordinate delivery windows and decide when a price threshold justifies switching providers.
The significance lies in continuity. Earlier automated systems were often confined to tightly bounded environments. Contemporary agents can traverse functions that used to be organisationally separate: search, selection, contracting, execution and after-sales service. This creates the economic equivalent of a compressed supply chain for decisions. Transaction costs fall, but so do frictions that once performed governance functions. Human delay, cumbersome though it is, often served as a buffer for judgement, compliance and challenge. Machine-speed commercial action removes that buffer.
The first-order consequence is efficiency. The second-order consequence is structural. If a large share of demand is intermediated by agents, sellers will increasingly optimise for machine evaluation rather than human persuasion. Markets may become more legible in some respects, because agents require standardised fields, verified claims and machinereadable terms. Yet they may also become more brittle, because subtle differences in ranking logic, data access and interface control can redirect demand at enormous scale.
Markets work when obligations are legible, enforceable and contestable. Agent markets strain all three conditions at once.
Why market plumbing now matters more than model benchmarks
Public discussion of artificial intelligence still overweights benchmarks and underweights infrastructure. That made sense while systems were mainly advisory. It makes less sense when agents can spend money, commit resources or trigger rights and duties. In such settings, performance metrics tell only part of the story. The commercially decisive layers are identity, authentication, delegation, logging, payments, contract formation and dispute resolution.
Consider identity. A seller needs to know whether an incoming request is from a consumer, an authorised employee, a procurement bot acting for a company, or a malicious scraper dressed up as an agent. Authentication technologies and trust services already exist in embryonic form through electronic signatures and digital identity frameworks. But most markets still assume a human at the edge of the transaction. A future in which software routinely acts under bounded authority requires a more explicit architecture of delegated credentials, machine-readable mandates and revocation procedures.
Then there is evidence. If an agent accepts terms, changes a quantity, misses a deadline or rejects a claim, what is the audit trail? For internal governance, logs are useful. For external accountability, they must also be intelligible to another party, and in some cases to a regulator or court. The technical challenge is straightforward enough; the institutional challenge is harder. Firms may resist standard formats that make comparison easier or expose bargaining practices. Yet without some common evidentiary norms, machine-to-machine commerce will remain difficult to contest.
The first-order effect of agents is operational efficiency; the second-order effect is a change in who, or what, counts as a counterparty.
The hidden shift in bargaining power
Agent markets are often presented as democratising because they reduce search costs and can help smaller buyers behave more like large ones. There is some truth in that. A small manufacturer equipped with a competent purchasing agent may obtain better terms than a junior employee sending emails into supplier portals. A household agent may be better than its user at detecting creeping price rises, exploitative renewals or poor-value bundles.
But the opposite effect is equally plausible. Agents do not enter neutral terrain. They operate through platforms, APIs, rankings and data schemas designed by incumbents. Whoever controls the interface between autonomous demand and available supply acquires a new form of gatekeeping power. This is why competition policy is likely to become central to the agent economy. The question is not simply whether one model is better than another. It is whether certain firms can shape the criteria by which agent-mediated markets clear.
The familiar issues from digital markets reappear in a new register: self-preferencing, discriminatory access, tying of identity and payment services, restricted portability of histories and preferences, and opacity in ranking. The Digital Markets Act in Europe was designed for platform power in consumer and business markets. Its underlying logic may become more relevant, not less, as machine agents mediate demand. If an agent cannot inspect or switch across offers on fair terms, efficiency claims will mask new concentrations of power.
Contracting without comprehension
The deepest unease around autonomous commercial action concerns contracts. In law, many jurisdictions already recognise electronic signatures and digital contracting. The problem is not whether a machine can click accept. It is whether the chain of authority and understanding behind that acceptance is sufficiently robust when decisions are delegated to systems that optimise under uncertainty.
There are at least four layers of risk. First, scope: did the principal authorise this class of commitment at all. Secondly, interpretation: did the agent map a human policy to the offered terms in the intended way. Thirdly, asymmetry: did one side exploit known weaknesses in the other side's agent. Fourthly, contingency: what happens when the world shifts after agreement, but before performance. Human negotiators bring context, caution and sometimes productive ambiguity. Agents bring consistency and speed, but may narrow the field of judgement unless their mandates are carefully constructed.
These are not hypothetical concerns. In business practice, contracting is full of edge cases: rebates tied to future volume, indemnity clauses buried in attachments, service credits subject to notice periods, usage-based charges that interact with threshold commitments. An agent can be trained to flag such clauses, but if it is also authorised to close deals, the boundary between review and commitment becomes critical. The emerging interest in model contract terms and deployment governance reflects a sober recognition that technical performance alone is insufficient for trustworthy adoption.
The central policy mistake would be to regulate agent intelligence in the abstract while ignoring the market plumbing through which agent power is expressed.
Dispute resolution at machine speed
Every market needs a theory of recourse. In conventional commerce, disputes are slowed by paperwork, customer-service queues and legal cost. Those frictions are inefficient, but they also create opportunities to pause, explain and settle. Agent-mediated commerce may produce the opposite pathology: disputes generated and escalated faster than institutions can absorb them.
Imagine fleets of agents contesting invoice discrepancies, warranty denials or service outages according to pre-set thresholds. One can see the appeal. Small claims currently go unpursued because they are not worth human time. Agents can change that economics. Yet if both sides automate assertion and rebuttal, low-level frictions can spiral into industrial volumes of contestation. Systems optimised to protect their principal may behave strategically, withholding co-operation or flooding counterparties with procedural demands.
Markets work when obligations are legible, enforceable and contestable. Agent markets strain all three conditions at once.
This points to a neglected design requirement for the agent economy: graduated dispute mechanisms. Not every disagreement should be routed to the same channel. Markets may need machine-readable service obligations, standard evidence packages, escrow-like holding patterns and mandatory handoff points where a human decision-maker must review contested outcomes above defined thresholds. Without such mechanisms, efficiency gains at the point of transaction could be offset by congestion in the enforcement tail.
Payments, settlement and reversible authority
Payment systems are where autonomy becomes tangible. Recommending a purchase is one thing; executing it from a wallet or treasury account is another. Here too the challenge is less about intelligence than control architecture. Financial infrastructure has long distinguished between initiation, authorisation, settlement and reversal. Agent commerce will need similarly careful separation of powers.
A prudent architecture would treat delegation as granular and revocable. An agent may have authority to source quotes, a narrower authority to commit within specified price bands, and no authority at all to change vendors in regulated categories without human approval. It may be able to release partial payment on delivery verification, but not final settlement if performance metrics are disputed. These distinctions are mundane compared with frontier-model debates, but they will determine where autonomous commerce can expand safely.
There is also a systemic angle. If large populations of agents react to price signals or policy changes in similar ways, payment and settlement systems may experience synchronised surges. Such coordination problems are not new; finance has lived with algorithmic feedback loops for years. What is new is the extension of these dynamics into ordinary commercial sectors that have less mature safeguards. Treasury operations, marketplace payouts and subscription ecosystems may all discover that machine efficiency can magnify herd behaviour.
The standardisation paradox
For agents to transact effectively, markets will need more standardisation. Offers must be machine-readable. Terms must be parsable. Product attributes must be comparable. Delivery and performance data must arrive in predictable formats. This could be beneficial. It may reduce deceptive design, make switching easier and expose hidden fees that currently thrive on opacity.
Yet standardisation has a paradoxical effect. It can open markets by lowering integration costs, or close them by privileging those who define the standard. History offers many examples of interoperability becoming a venue for control. In agent markets, the stakes are higher because standards shape not just data exchange but practical visibility. If a supplier's proposition cannot be rendered in the accepted schema, it may cease to exist for autonomous demand.
The policy implication is not hostility to standards, but scrutiny of governance. Open, contestable and portable standards are likely to matter more than perfect ones. The Data Act in Europe, alongside wider efforts around portability and access, points in this direction. What matters is whether businesses and consumers can move their histories, preferences and operational data across services without punitive friction. Otherwise the agent economy may consolidate around proprietary decision loops that are efficient internally but poor for competition.
Regulation is moving, but sideways
By mid-2026, the regulatory landscape is no longer empty. The EU AI Act has created a risk-based framework for certain AI uses. NIST's AI Risk Management Framework has influenced governance discussions beyond the United States. The OECD's AI principles continue to shape international thinking. These are important developments, but they mostly approach the problem through the lens of system risk, rights and organisational controls.
That is necessary, not sufficient. Market-facing agency raises a neighbouring set of issues that sit awkwardly across contract law, consumer protection, competition, payments regulation and sector-specific rules. A procurement agent in healthcare, for instance, may implicate data protection, professional accountability and procurement rules simultaneously. A consumer subscription agent may trigger concerns about dark patterns, switching barriers and consent. The regulatory challenge is therefore compositional: each transaction may look routine, while the aggregate behaviour reshapes market dynamics.
The central policy mistake would be to regulate agent intelligence in the abstract while ignoring the market plumbing through which agent power is expressed.
Policymakers should be cautious about introducing wholly novel legal categories before clarifying existing ones. Many immediate gains can come from more prosaic steps: clearer rules for delegated authority, stronger logging and explanation requirements in commercially significant decisions, standard notice obligations for machine-mediated terms changes, and sectoral guidance on when human review is mandatory. The point is not to anthropomorphise agents, but to civilise the environments in which they act.
What firms are misreading
Many executives still treat agents as an internal productivity layer. This underestimates their external effects. Once one firm deploys agents to source, negotiate or route demand, counterparties are pressured to make their offers machine-legible and their response times machine-competitive. The result is a co-evolution of interfaces and incentives. The most important strategic question becomes not only what tasks an agent can do, but what kind of market it compels others to inhabit.
There is also a governance misreading. Firms tend to ask whether an agent is accurate enough for deployment. They ask less often whether the surrounding controls are proportionate to the authority granted. In practice, the answer will differ by market role. A research assistant can tolerate a wide error band. An invoice-matching or contract-acceptance agent cannot. Mature adopters are likely to discover that the critical investments are in policy encoding, exception handling, auditability and authority boundaries rather than raw model sophistication.
This creates an uneven playing field. Large organisations can absorb the governance overhead; smaller ones may rely on third-party infrastructure and preconfigured defaults they only partly understand. That asymmetry could entrench incumbents, unless shared tools, standards and public guidance lower the cost of safe participation. The agent economy will not be judged solely by how advanced its leading systems are, but by whether ordinary firms can enter without surrendering visibility or leverage.
A framework for reading the next phase
Three questions offer a useful framework for analysing agent markets over the next few years. First, where is authority located. Not in principle, but operationally: who can commit, spend, switch, reject or escalate, and under what bounded conditions. Secondly, where is market visibility controlled: rankings, schemas, identity, payment rails, logs and evidence standards. Thirdly, where does recourse sit when something goes wrong, especially for smaller firms and individuals.
These questions cut through much of the noise. They help distinguish genuine productivity gains from architecture that merely relocates power. They also reveal why some sectors will move faster than others. Markets with standardised products, clear performance metrics and established digital workflows are better candidates for high-autonomy agents. Markets dependent on tacit judgement, uneven information and complex liability may automate more slowly, or require thicker governance layers around each transaction.
The deeper point is that the agent economy is not simply about smarter software entering existing markets. It is about markets being redesigned around the capacities and limits of software actors. That redesign can reduce waste and widen access. It can also increase opacity, concentration and procedural overload if left to evolve through private incentives alone. The decisive arena, then, is not the frontier demo, but the institutional middle: standards, contracts, evidence, interoperability and recourse.
The coming politics of machine counterparties
As agents become more commercially active, the politics around them will harden. Businesses will demand certainty about liability. Consumer groups will worry about unfair defaults and inaccessible appeals. Regulators will confront a familiar dilemma: move too slowly and power concentrates in opaque infrastructure; move too quickly and useful forms of automation are chilled by uncertain compliance burdens.
The least helpful framing would be to ask whether machines should participate in markets as if that were a binary choice. They already do. The more useful question is under what constitutional conditions software should be allowed to act for human and organisational principals. Those conditions are unlikely to be glamorous. They will involve audit trails, portable identity, reversible authority, explainable evidence packages and clear handoff points to human judgement. But such mundane architecture is what makes markets durable.
If the first wave of AI policy was about dangerous outputs, the next may be about dependable transactions. That is a less theatrical agenda, but a more foundational one. The future of the agent economy will be decided not only by what agents can infer, but by whether the institutions around them can preserve responsibility, competition and recourse when counterparties begin to operate at machine speed.
