The economy is acquiring a new kind of participant
For two decades, software has steadily absorbed clerical work, recommendation, search and routine decision support. What is changing now is not merely that systems can generate language or code, but that they can increasingly take actions across digital environments: compare options, compose requests, trigger workflows, monitor outcomes and adapt their next step. In economic terms, this moves software from being a passive tool towards something closer to an operational actor.
The phrase “agent economy” captures this transition. It refers to markets and organisations in which autonomous or semi-autonomous software systems perform tasks that once required a human intermediary, and in some cases interact directly with one another. These systems do not need consciousness to matter economically. They need only enough competence, memory, connectivity and delegated authority to make decisions that have consequences.
That may sound futuristic, but the ingredients are already visible. Digital platforms have long used automated bidding, dynamic pricing, fraud detection and supply-chain optimisation. What is new is the convergence of large language models, tool use, workflow orchestration and API-rich business infrastructure. Together, these make it easier to build systems that can handle unstructured requests, translate them into actions and carry them through across multiple steps.
The decisive shift is not that software can speak more naturally, but that it can increasingly do things on someone’s behalf.
This distinction matters. A chatbot that drafts an email is useful. An agent that reads an inbox, prioritises requests, queries internal systems, negotiates calendar trade-offs and closes a loop with suppliers or customers occupies a different economic category. It begins to substitute not just for isolated tasks but for coordination itself.
From automation to delegated action
Traditional automation works best where processes are stable, inputs are structured and outputs can be tightly specified. The industrial logic is familiar: codify a workflow, remove variance, increase throughput. Agentic systems extend this logic into messier terrain. They can parse ambiguous instructions, generate intermediate plans and recover from minor surprises. In effect, they lower the cost of handling exceptions.
This has important implications for transaction costs, a concept associated with Ronald Coase and later Oliver Williamson. Firms exist in part because using the market is costly: one must search for partners, draft agreements, monitor performance and resolve disputes. If software agents can reduce these costs at scale, some activities that are now internal to firms may become easier to contract out, while others may be re-centralised because coordination inside the firm becomes dramatically cheaper.
In practice, the first wave is likely to be hybrid rather than fully autonomous. Humans will define objectives, policy boundaries and escalation rules. Agents will handle discovery, synthesis, scheduling, procurement support, documentation and repetitive communications. Yet even this modest degree of delegated action can alter organisational economics. Middle layers devoted to routing information, chasing approvals and managing routine exceptions may shrink or be redefined.
The result is not a world without management, but one in which management becomes more about setting constraints, validating outputs and auditing decisions than manually moving work from one desk to another.
Why trust, not intelligence alone, is the binding constraint
Much discussion of advanced software still focuses on benchmark performance. But economic adoption depends on something more prosaic: trust. Firms do not deploy systems into sensitive workflows because they produce eloquent answers. They do so because the systems are reliable enough, observable enough and governable enough for specific contexts.
The relevant question is therefore not, “How clever is the model?” but, “Under what conditions can it act?” In payments, healthcare, hiring, insurance, logistics and public administration, the answer turns on traceability, permissions, identity, security and liability. Even a highly capable system becomes commercially constrained if no one can determine why it acted, whose data it accessed or who is responsible when it fails.
The decisive shift is not that software can speak more naturally, but that it can increasingly do things on someone’s behalf.
The research literature has increasingly reflected this concern. The National Institute of Standards and Technology’s AI Risk Management Framework emphasises governance, mapping, measurement and management of AI risks. Meanwhile, work from international bodies such as the OECD has stressed robustness, transparency and accountability as preconditions for trustworthy deployment. These are not abstract principles. They are design requirements for any economy in which software agents are expected to carry authority.
In the agent economy, capability opens the door, but accountability decides which rooms software is allowed to enter.
This is why the next phase of competition may centre less on model size and more on infrastructure for verification: audit trails, policy engines, rights management, sandboxing and secure identity layers. The firms that can make agents legible to risk, compliance and procurement functions will enjoy a structural advantage over those that merely promise autonomy.
Labour markets will change through task reassembly
Predictions about employment effects often swing between alarm and complacency. Neither is particularly useful. A better frame comes from task-based analysis, including work by the OECD and the International Labour Organization, which examines how technologies alter bundles of tasks rather than erase occupations wholesale. Most jobs combine routine, social, physical and judgement-based elements. Agents will unbundle these mixtures unevenly.
Clerical, administrative and coordination-heavy roles are especially exposed because so much of their value lies in moving information across systems, checking status, drafting standard communications and maintaining procedural flow. That does not imply immediate disappearance. It does suggest a steady transfer of routine cognitive labour from people to software, with remaining human work concentrated in oversight, exception handling, relationship management and contextual judgement.
At the same time, new work will emerge around supervising agents, designing workflows, defining escalation policies, curating knowledge sources and evaluating performance. The pattern may resemble earlier waves of enterprise software adoption: fewer people doing repetitive process execution, more people doing systems stewardship and cross-functional problem-solving.
Still, adjustment costs could be significant. Workers whose skills were built around coordination rather than deep domain expertise may face pressure first. This raises familiar policy questions about retraining, credential portability and labour-market resilience. It also raises a newer question: how should productivity gains from delegated software be shared when they accrue primarily to owners of data, distribution and compute rather than to the employees whose tasks are being codified?
Markets may become faster, thinner and more machine-readable
As agents transact more often, markets may evolve to suit them. Human markets rely on persuasion, branding, relationships and a degree of tolerated ambiguity. Machine-mediated markets reward structured information, standardised interfaces and explicit constraints. In many sectors, that will favour offers that can be compared automatically: prices, service levels, delivery windows, provenance records, contract terms and performance metrics expressed in machine-readable formats.
This could make some markets more efficient. Search frictions may fall. Price discovery may improve. Small suppliers with clean data and interoperable systems may gain access to buyers that were once too costly to reach. Yet efficiency can come with concentration risks. If a few platforms, protocols or identity systems become the de facto channels through which agents discover and transact, market power may shift towards whoever controls those gateways.
The history of digital advertising, app distribution and e-commerce suggests that lower transaction costs do not automatically produce open competition. They can also entrench bottlenecks. The European Commission, among other regulators, has repeatedly argued that digital markets can tip towards gatekeepers when network effects, data advantages and switching costs reinforce one another. Agent-mediated commerce may reproduce this pattern unless interoperability and contestability are built in early.
There is also the possibility of machine collusion or tacit coordination. The UK’s Competition and Markets Authority and other authorities have long studied how pricing algorithms can facilitate anti-competitive outcomes even without explicit human conspiracy. As autonomous systems gain greater discretion, antitrust analysis will need to examine not only code but also emergent behaviour in multi-agent environments.
Contracts are becoming operational rather than static
One underappreciated feature of the agent economy is its effect on contracting. Conventional contracts are documents interpreted by people, enforced by institutions and revisited episodically. Agentic systems push towards more operational forms: agreements whose terms are partially embedded in workflows, permissions and automated checks.
In the agent economy, capability opens the door, but accountability decides which rooms software is allowed to enter.
This does not mean legal prose disappears. It means that execution increasingly depends on technical controls as much as on after-the-fact interpretation. Spending limits can be encoded. Approval chains can be formalised. Data access can be scoped. Service conditions can trigger alerts or automatic remediation. In this sense, part of the contract becomes a living control system.
That shift could reduce some forms of opportunism and administrative overhead. But it also creates a new governance challenge: policy translation. Organisations must convert broad legal and ethical principles into precise machine-operable rules, then continuously update them as regulations and business conditions change. Anyone who has worked with real contracts knows how difficult that is. Ambiguity is not a bug in many agreements; it is a feature that allows flexibility. Agents are less comfortable with ambiguity.
The likely result is a layered model. High-level terms remain human-readable and subject to conventional law. Day-to-day execution becomes increasingly automated through policy tooling, system permissions and event monitoring. Economic value will accrue to those able to bridge these layers without losing nuance or control.
Governance must move closer to the point of action
Many organisations still govern software through periodic review: approve a system, deploy it, audit it later. That is poorly suited to agents operating in dynamic environments. If software can take thousands of small actions a day, governance cannot be a once-a-quarter ritual. It must be woven into runtime.
This implies a shift from model-centric governance to action-centric governance. The core concern is less the existence of an AI system than the permissions attached to it, the resources it can access, the thresholds for intervention and the records it leaves behind. A low-risk summarisation tool and an identically capable system authorised to purchase inventory or alter customer records belong in entirely different risk categories.
Public frameworks are beginning to reflect this. The European Union’s AI Act adopts a risk-based approach, while guidance from standards bodies increasingly stresses documentation, monitoring and human oversight. For firms, however, compliance will not be enough. The operational challenge is to create governance architectures that are fast enough for deployment teams, clear enough for auditors and resilient enough for regulators.
The most important control in an agentic system may not be what the model knows, but what the organisation permits it to touch.
This is likely to elevate the importance of identity management, access control, logging and incident response. In effect, the institutions of cybersecurity and enterprise risk are converging with the institutions of AI deployment. The agent economy will mature not when models become magical, but when operating controls become mundane and dependable.
Public services and infrastructure will face a hard test
The attraction of agents in the private sector is clear: lower administrative costs, faster response times and round-the-clock handling of routine work. Public services have similar incentives, perhaps more so, because they are often burdened by backlogs, fragmented systems and rigid processes. Yet the public realm poses harder legitimacy questions.
When software helps determine eligibility, prioritise inspections, flag fraud or route vulnerable citizens through service pathways, errors carry civic consequences. A mistaken purchase order in a business may be inconvenient. A mistaken benefits decision or health triage recommendation may be deeply harmful. Public institutions therefore need stronger standards of explainability, appeal and procedural fairness than many commercial contexts demand.
Here the lesson from earlier digital government efforts is instructive. Administrative simplification is valuable, but only if citizens can understand decisions, contest them and reach a human when needed. The more capable agents become, the stronger the temptation to let them absorb scarce administrative labour. The danger is not only technical failure, but bureaucratic opacity at scale.
Well-designed public adoption would focus first on bounded support roles: document handling, appointment management, internal drafting and case preparation, with clear escalation paths and transparent records. Infrastructural modernisation, not just model deployment, will determine whether public-sector use becomes a gain in state capacity or another chapter in digital mistrust.
The most important control in an agentic system may not be what the model knows, but what the organisation permits it to touch.
Global rules will matter as much as local innovation
The agent economy will not develop in a legal vacuum. Rules on data protection, consumer rights, financial supervision, competition and employment law will shape where autonomous systems can be used and how quickly they spread. Because digital systems cross borders more easily than legal concepts do, fragmentation is likely.
Europe’s regulatory approach emphasises precaution, rights and risk categorisation. The United States has tended to move through sector-specific enforcement, standards development and state-level experimentation. International organisations such as the OECD and IMF have focused on policy coordination and macroeconomic implications. This divergence will influence business design. Some organisations will build to the strictest common denominator; others will segment capabilities by jurisdiction.
There is a broader geopolitical dimension too. Countries that provide trusted digital identity, robust payment rails, clear liability norms and interoperable public standards may become more fertile ground for agent-based commerce. In that sense, the agent economy is not merely a software race. It is also a contest in institutional engineering.
That institutional layer is often overlooked because it lacks glamour. Yet markets run on confidence as much as invention. Just as containerisation transformed trade through standards rather than spectacle, the spread of software agents may depend less on breakthroughs in language generation than on boring but decisive advances in verification, dispute resolution and cross-system interoperability.
The firms that benefit most may be the ones that redesign themselves
There is a temptation to treat agents as another application to be bolted onto existing workflows. Some gains will indeed come that way. But the larger returns are likely to accrue to organisations willing to redesign processes around delegated software from the outset.
That means simplifying approval chains, cleaning underlying data, defining clear ownership of decisions and identifying where human judgement genuinely adds value. It also means resisting the common mistake of automating chaos. An agent placed atop contradictory policies, fragmented systems and poor incentives will not create order; it will reproduce confusion faster.
Organisational redesign could alter corporate boundaries as well. Smaller teams may manage larger operational scopes. Specialist expertise may become more leveraged if agents handle routine preparation and follow-through. The premium may rise on people who can combine domain knowledge, process thinking and governance literacy.
In this sense, the agent economy is not simply about substituting labour. It is about recomposing the firm. Some hierarchies will flatten. Some support functions will contract. New control functions will expand. The balance between centralisation and decentralisation may shift repeatedly as firms learn where autonomy works and where it creates unacceptable risk.
An economy of agents will still be governed by human choices
It is tempting to discuss autonomous software as though it follows an inevitable path from novelty to ubiquity. History suggests otherwise. Technologies spread when they fit institutions, incentives and social expectations. Where they conflict with law, legitimacy or practical workflow, adoption stalls or retreats.
The agent economy will therefore be built through choices about delegation. Who gets to authorise software to act? What records must be kept? Which domains require human review? How are losses allocated when systems err? Which interfaces remain open, and which become controlled chokepoints? These are economic questions because they influence competition, productivity and bargaining power. They are political questions because they distribute agency itself.
That is why the most serious debate is not about whether software will appear more agentic. It will. The real debate concerns the terms under which agency is lent to machines. If those terms are clear, contestable and well governed, autonomous systems could lower administrative drag and widen access to sophisticated services. If they are opaque or overly concentrated, the same systems could deepen dependency and blur responsibility.
The agent economy, then, is best understood not as a science-fiction threshold but as an institutional project. Software is learning to act. The task now is to decide, with unusual care, where action should be allowed to begin and where human judgement must remain firmly in charge.
