From tools to actors
The phrase “agent economy” captures a simple but consequential shift: software is moving from being a passive instrument used by people to an active participant that can complete goals on their behalf. In practical terms, an agent can interpret instructions, gather information, choose among options, trigger workflows and, in some cases, transact across digital systems. That does not make software human-like. It means, rather, that a growing share of economic activity can be organised through semi-autonomous computational entities.
This shift has a long prehistory. Long before today’s language-model systems, economists and computer scientists were considering how software agents might search markets, coordinate tasks and reduce transaction costs. What is new is the combination of broader technical capability, cheaper computation, ubiquitous APIs, cloud infrastructure and digital commerce. Together, these conditions have made agents usable at scale rather than merely interesting in theory.
The agent economy is not a bolt from the blue; it is the meeting point of automation, online markets and machine intelligence.
A timeline helps clarify what changed, when, and why it matters. The story begins not with generative AI, but with earlier efforts to digitise coordination itself.
1960s–1980s: the intellectual foundations
The conceptual roots of the agent economy lie in several older traditions. One is the economics of transaction costs, associated with Ronald Coase and later Oliver Williamson, which asked why firms exist and how coordination costs shape organisational form. If the cost of finding information, comparing options and enforcing decisions falls, the boundary between internal management and external markets can shift. Digital systems promised precisely that.
Another foundation came from artificial intelligence research into planning, expert systems and distributed problem-solving. By the 1980s, researchers were already exploring how independent software entities might cooperate or compete in “multi-agent systems”. These were often laboratory models, but they posed questions that now feel strikingly current: how should agents divide labour, communicate intent and resolve conflict?
At the same time, enterprise computing began to formalise business processes. Databases, electronic data interchange and early workflow systems converted previously informal activities into machine-readable steps. This mattered because one cannot automate what has not first been structured. The eventual rise of agents depended on years of painstaking digitisation that made inventory, logistics, procurement and customer records available to software.
1990s: the web creates an open commercial substrate
The commercial internet changed the landscape by linking information, users and merchants on a common network. Search engines reduced the cost of discovery. E-commerce standardised online catalogues and ordering. Digital advertising connected attention to measurable outcomes. In parallel, online payment systems and card networks made remote transactions routine.
Scholars noticed early that software could thrive in such an environment. A widely cited paper by Pattie Maes, Robert Guttman and Alexandros Moukas in 1999 described “agents that buy and sell”, arguing that autonomous software could help users navigate information overload and market complexity. Their analysis now reads as an early map of agent commerce: recommendation, comparison shopping, negotiation and delegated purchasing.
The agent economy is not a bolt from the blue; it is the meeting point of automation, online markets and machine intelligence.
Yet the infrastructure was still immature. Most websites were designed for human browsing, not machine-to-machine action. Data formats were inconsistent, APIs were limited and trust remained a barrier. The 1990s established the internet as a marketplace, but not yet as a habitat for broadly capable economic agents.
The internet made markets searchable; later infrastructure would make them machine-operable.
2000s: platforms, APIs and the routinisation of digital work
The 2000s were less dramatic in rhetoric than the present moment, but crucial in substance. Firms standardised digital processes, while web services matured into programmable infrastructure. Application programming interfaces allowed software to access payments, maps, messaging, identity services and cloud storage without rebuilding everything from scratch. This was a turning point: software no longer needed merely to display information; it could act across modular services.
In business operations, robotic process automation and workflow software began targeting repetitive office tasks. These systems were brittle by current standards, but they demonstrated an economic appetite for delegating structured work to machines. At the same time, online labour platforms, marketplaces and app ecosystems expanded the geography of work and reduced the frictions of matching supply with demand.
Cloud computing further changed the cost structure. Instead of large upfront investment in servers and maintenance, firms could rent computation as needed. That lowered the barrier to experimentation and made software-intensive business models more scalable. The seeds of the agent economy were being sown in the mundane language of integration, orchestration and service abstraction.
2010s: machine learning improves judgement at scale
The next decisive layer was machine learning. As deep learning advanced, software became better at pattern recognition in language, images, demand forecasting, anomaly detection and recommendation. This did not yet create general-purpose agents, but it improved one of the central ingredients of agency: the ability to make useful inferences under uncertainty.
Economic institutions took notice. Research from the Bank for International Settlements and others began tracking how digitalisation, intangibles and data-driven business models were altering market structure and productivity dynamics. Meanwhile, the OECD documented the expansion of digital trade and the rising value of data as an economic resource. Software was becoming not only a tool of production but also a mechanism of coordination.
The decade also normalised digital assistants and conversational interfaces, though their capabilities were limited. Users became more comfortable giving software natural-language instructions, while firms grew accustomed to AI systems embedded in ordinary workflows. Importantly, the broader economy was now rich in APIs, cloud services and digital records. By the late 2010s, the missing piece was a more flexible interface between human goals and machine execution.
2020–2022: foundation models change the interface
Large language models transformed this interface by making software more adept at interpreting messy instructions and generating structured outputs. For the first time, a single model could summarise documents, draft text, extract fields, write code, classify requests and decide which tool to call next. That versatility made the notion of a software “agent” newly practical.
What changed economically was not intelligence in the abstract, but the cost of orchestration. Tasks that once required a custom rule set for every workflow could now be handled through natural-language prompts, retrieval from external data sources and tool use. McKinsey and the International Monetary Fund both argued in 2023 and 2024 that generative AI could affect a large share of work activities, especially in knowledge-heavy sectors. Their estimates vary, but the common point is clear: language became computable in a commercially useful way.
The internet made markets searchable; later infrastructure would make them machine-operable.
Still, these systems were not fully autonomous. They could be inaccurate, overconfident and difficult to govern. But they lowered the threshold for building software that appeared goal-directed. That encouraged a wave of experiments in research assistance, customer support, software development, internal operations and digital procurement.
2023: the vocabulary of agents enters the mainstream
In 2023, “agents” became a mainstream business and technical term. Developers began linking language models to external tools such as search, browsers, databases and productivity software. Research communities explored benchmarks for tool use, planning and long-horizon tasks. Policy institutions, meanwhile, started distinguishing between narrow AI features and systems that could execute sequences of actions with limited supervision.
This was also the year when the public debate sharpened. Advocates emphasised productivity gains and the possibility of software handling routine cognitive work. Sceptics pointed to reliability, security and accountability. Both sides had a point. Agents looked less like magical intelligence than like probabilistic workflow engines: powerful in bounded settings, hazardous in poorly designed ones.
Even so, the direction of travel was unmistakable. Once software could parse intent, consult context and take action through APIs, it began to resemble a market participant. It could source options, compare prices, generate a recommendation and submit the chosen action into another system. The line between assistance and delegation started to blur.
Once software can interpret intent and act through digital rails, it starts to look less like a tool and more like a participant in the market.
2024: from demos to operational questions
By 2024, the central question was no longer whether agents were possible, but where they could be deployed responsibly. Attention shifted from eye-catching demonstrations to the operational plumbing of the agent economy: identity, permissions, audit trails, retrieval quality, observability, model evaluation and fallback mechanisms. Firms discovered that the hard part was not generating fluent text; it was making action dependable.
Regulators and standard-setters were moving as well. The European Union adopted the AI Act, offering one of the first broad legal frameworks for AI systems, especially those used in high-risk settings. NIST in the United States advanced its AI Risk Management Framework, emphasising governance, validity, security and accountability. These efforts did not target the agent economy alone, but they addressed the conditions under which autonomous or semi-autonomous systems might be trusted in practice.
Meanwhile, central banks and international organisations were increasingly focused on digital money, tokenisation and faster payment systems. Not all of this is directly about AI agents. But if software is to transact at scale, payment and settlement rails must become more programmable and interoperable. The agent economy depends as much on institutional plumbing as on model capability.
The hidden enablers: identity, payments and trust
For software agents to play a larger economic role, three enabling layers matter more than headlines often suggest. The first is identity. An agent needs a recognised way to authenticate itself, inherit permissions and operate within clearly defined boundaries. Without that, autonomous action is either impossible or too risky to permit.
The second layer is payment and settlement. Economic agency requires the ability not just to decide, but to commit resources. That may involve procurement thresholds, escrow arrangements, internal budgets or external payment rails. Research from the Bank of England, the BIS and others on digital money and tokenisation suggests that more programmable financial infrastructure could eventually reduce frictions in machine-mediated exchange.
Once software can interpret intent and act through digital rails, it starts to look less like a tool and more like a participant in the market.
The third layer is trust and auditability. Human institutions tolerate mistakes from employees because accountability structures exist around them. Software agents require their own equivalents: logs, explainability where possible, access controls, policy constraints and mechanisms for redress. In this sense, the agent economy is not simply a technological challenge. It is a governance problem embedded in commercial systems.
What changes inside firms
The most immediate effects of the agent economy are likely to appear within organisations before they transform open consumer markets. Firms already possess the ingredients agents need: proprietary data, defined workflows, permissioned environments and measurable tasks. Agents can therefore be deployed first in areas such as procurement support, compliance triage, contract review, internal help desks, software maintenance and sales operations.
This matters for the theory of the firm. If agents reduce the cost of coordination, they may allow organisations to operate with fewer managerial layers in some functions while increasing the importance of oversight in others. Some work will be automated outright; some will be re-bundled into supervisory roles; some will become more valuable because humans remain responsible for exception handling, judgement and institutional legitimacy.
Productivity effects are plausible but uneven. The IMF has argued that AI may complement high-skilled labour while exposing other tasks to displacement. The OECD has likewise stressed that policy choices and workplace design will shape whether gains are broad-based or concentrated. In the agent economy, the unit of analysis is often not the job but the workflow. That makes change both more granular and more pervasive.
What changes in markets
As agents become more capable, market structure could shift in subtle ways. Search and discovery may move away from users manually comparing options towards agents pre-filtering choices. Negotiation may become more data-driven and continuous. Procurement could increasingly occur between software systems rather than through human correspondence. In digital services especially, machine-readable quality signals may matter more than brand messaging or interface polish.
There are, however, reasons for caution. Agents may reinforce incumbency if only the largest firms can provide the data, distribution and compute required to train and operate them effectively. They may also create new bottlenecks around identity, access and platform rules. Automated interactions can produce speed and efficiency, but also strategic manipulation, collusion risks or cascading errors if many systems optimise in similar ways.
Competition authorities and regulators will therefore face unfamiliar questions. How should consumer protection work when software is acting as an intermediary? Who is liable when one agent transacts with another under delegated authority? How should market abuse be detected when decisions are distributed across automated systems? These are not hypothetical puzzles for the distant future; they are emerging design constraints for digital commerce now.
The next phase
The agent economy remains early, but its trajectory is becoming clearer. The most durable progress is likely to come not from systems that promise unrestricted autonomy, but from narrower agents embedded in institutional settings with clear goals, high-quality data and robust controls. In that sense, the future may look less revolutionary than cumulative: more software-mediated decisions, more machine-to-machine coordination and more markets designed for computational participants.
That should not be mistaken for trivial change. Over time, small reductions in the cost of searching, deciding and executing can reshape how firms organise work and how markets clear. The significance of the agent economy lies precisely there. It is not merely about clever interfaces or chat-based assistance. It is about whether software can become a dependable counterparty in the economy’s routine transactions.
The timeline suggests a sober conclusion. Agents became possible because the internet made markets digital, cloud computing made infrastructure elastic, APIs made services callable and modern AI made language operational. The next chapter will depend less on spectacle than on institutions: standards, regulation, payments, identity and governance. Economies change when coordination changes. The rise of agents is one more chapter in that longer history.
