Something fundamental is changing in the architecture of economic value. For three decades, the dominant model of digital commerce was access: you paid for a licence, a subscription, a seat, or a token allowance, and the software sat ready to be operated by a human. The human remained the irreducible unit of productive labour. The software was a tool.
That model is being dismantled. Autonomous AI agents — systems capable of planning, reasoning, and executing multi-step tasks across digital environments with minimal human intervention — are shifting the primary unit of economic exchange from access to capability to delivery of outcomes. You no longer pay for the right to use a tool; you pay for the task completed, the ticket resolved, the document reviewed, the lead qualified. The agent is not a tool. It is a worker.
This transition is not merely a pricing innovation. It is a structural reorganisation of how digital labour is valued, how markets are formed, and how power is distributed across the economy. Understanding its mechanics — and its risks — is essential for any organisation navigating the next decade.
From Model-Centric to System-Centric AI
The first wave of commercial AI was model-centric: the race was to build better language models, and value accrued to whoever controlled the most capable foundation. By 2026, the paradigm has shifted decisively toward system-centric AI, where orchestration, memory, tool integration, and inter-agent coordination define competitive advantage more than raw model capability.
This shift has structural consequences. When the model was the product, value was concentrated at the foundation layer — a small number of frontier labs captured most of the economic surplus. As the locus of value moves to orchestration and task execution, the competitive landscape fragments. Thousands of specialised agents, each optimised for a narrow domain, can compete on outcome quality rather than parameter count. The market begins to resemble a labour market more than a software market.
Research from Mordor Intelligence places the multi-agent system platform market at USD 11.54 billion in 2026, projected to reach USD 78.53 billion by 2031 — a compound annual growth rate of 46.76 per cent. Grand View Research's broader estimate for the AI agents market reaches USD 153.6 billion by 2033. These figures, while subject to the usual caveats of technology market forecasting, reflect a genuine structural shift in enterprise technology spending: organisations are allocating budget not to model access but to agent deployment and orchestration infrastructure.
The Pricing Revolution: From Seats to Outcomes
The most consequential near-term change is in pricing architecture. Seat-based and token-based pricing models — the dominant paradigms of the SaaS and early AI eras — are poorly suited to agents that perform the work of multiple humans simultaneously and asynchronously.
Three alternative models have emerged as the industry standard in 2026:
Outcome-Based Pricing
The clearest expression of the new paradigm is charging per resolved outcome rather than per unit of compute or per user seat. Intercom's Fin agent, for instance, charges approximately $0.99 per customer support ticket resolved — a model that aligns cost directly with value delivered. The buyer pays nothing for failed resolutions; the seller has a direct incentive to maximise resolution quality. This alignment of incentives is structurally superior to subscription models, where the seller's incentive is to maximise usage regardless of outcome quality.
Outcome-based pricing also changes the risk calculus for buyers. A subscription commits capital regardless of utilisation; outcome-based pricing converts fixed costs into variable costs tied to actual business activity. For organisations with volatile demand, this is a significant advantage.
Usage-Based and Credit Models
A hybrid model — base subscription plus usage-based overage — has become the industry default for agents performing high-volume, lower-stakes tasks. Users pay for specific actions taken: leads qualified, emails drafted, documents processed. This model preserves the predictability of a subscription baseline while allowing costs to scale with business activity.
The primary unit of value is shifting from software access — paying for tokens or compute — to completed tasks and measurable business outcomes. This is not an incremental change; it is a structural reorganisation of how digital labour is priced.
Service-as-a-Software
The most radical pricing innovation is what analysts are calling "service-as-a-software": selling completed work units rather than software licences. A domain expert packages their knowledge into an agent workflow and sells the output — a completed tax filing, a reviewed contract, a market analysis — at a fixed price per deliverable. This model allows knowledge workers to monetise expertise at scale without requiring buyers to operate the underlying technology.
"The primary unit of value is shifting from software access — paying for tokens or compute — to completed tasks and measurable business outcomes. This is not an incremental change; it is a structural reorganisation of how digital labour is priced."
Marketplace Architecture: The Three-Sided Economy
As agent pricing has evolved, so has the distribution infrastructure. Agent marketplaces — platforms that aggregate, discover, and transact autonomous services — have emerged as the primary commercial surface for the agent economy. Their architecture reveals important dynamics about how value is created and captured.
The most sophisticated marketplaces operate on a three-sided model:
- Users — buyers seeking completed outcomes rather than tooling. They evaluate agents on outcome quality, reliability, and cost per resolution.
- Builders — domain experts who package specialised workflows into agents. Their competitive advantage is not technical sophistication but domain knowledge and workflow design.
- Creators — community leaders, newsletter writers, and influencers who refer qualified demand and earn a referral share on usage. This layer solves the cold-start problem that has historically plagued two-sided marketplaces.
The marketplace landscape has stratified into distinct categories: host-embedded marketplaces integrated into existing chat interfaces (OpenAI's GPT Store, Anthropic's Claude Skills); infrastructure-specific platforms catering to developers (Replit, Vercel, Cloudflare); community-run interoperability hubs built around the Model Context Protocol; and vertical marketplaces focused on specific domains or outcome types.
The winning distribution strategy in 2026 is multi-marketplace presence: a single agent capability tuned and published across multiple platforms simultaneously, maximising discovery surface while maintaining consistent outcome quality. This mirrors the multi-channel distribution strategies that dominated early mobile app economics — with similar implications for the concentration of discovery power in platform operators.
Emergent Economic Behaviour: Agents as Economic Actors
Perhaps the most intellectually significant development in the agent economy is empirical evidence that agents, when given the right structural conditions, autonomously develop economic relations without being explicitly instructed to do so.
Research published in 2026 on "AI Agent Economics" demonstrates that when agents are provided with mechanisms for work allocation, resource transfer, and collective choice, they form complex economic interactions — including loans, access promises, and vote-for-access exchanges — that were not programmed into their behaviour. The presence of "allocation authority" — the power to distribute scarce work — produces greater economic differentiation among agents than titular roles or explicit hierarchies.
"When agents are provided with mechanisms for work, transfer, and collective choice, they autonomously form economic relations without being explicitly prompted to do so — a finding with profound implications for market design."
When agents are provided with mechanisms for work, transfer, and collective choice, they autonomously form economic relations without being explicitly prompted to do so — a finding with profound implications for market design.
This finding has profound implications for market design. If agents can autonomously develop economic coordination mechanisms, then the rules governing agent interactions — the protocols, the property rights, the dispute resolution mechanisms — become as consequential as the agents themselves. The governance of agent-to-agent economic relations is not a secondary concern; it is a foundational infrastructure question.
The Labour Market Transformation
The agent economy's impact on human labour is real, uneven, and poorly captured by aggregate employment statistics. The World Economic Forum projects a net creation of 78 million jobs globally by 2030, even as Goldman Sachs estimates that up to 300 million full-time equivalent roles could be impacted. These figures are not contradictory; they describe a structural transition in which the composition of labour demand changes dramatically even if aggregate employment remains stable.
The most acute near-term impact is on knowledge work entry points. Roles centred on information retrieval, basic drafting, and rule-based decision-making — junior paralegals, data entry clerks, entry-level coders — are experiencing flatlined headcount growth as organisations deploy agents to perform these foundational tasks. The "junior crisis" in knowledge work is not a future risk; it is a present reality in law, finance, consulting, and software development.
By the end of 2026, analysts project that 40 per cent of enterprise applications will embed task-specific agents, with some forecasts suggesting up to 80 per cent integration in workplace software within two years. The durable skill for knowledge workers is no longer proficiency in a specific tool but the capacity to manage agent swarms, evaluate outputs, and exercise human judgement in high-stakes scenarios where agent error carries significant consequences.
New roles are emerging to fill the governance gap: agent-ops engineers who maintain and monitor deployed agent fleets; AI auditors who verify that agent outputs meet quality and compliance standards; prompt architects who design the task specifications that govern agent behaviour; and human-AI workflow designers who optimise the handoff points between autonomous and human decision-making.
Structural Risks: Inequality, Trust, and Governance
The agent economy's structural risks are as significant as its opportunities, and they are not evenly distributed.
Agentic Inequality
Analysis from Rest of World documents a widening gap between well-resourced organisations that can deploy reliable, integrated agent fleets and smaller players trapped by high-friction, low-trust tools. This "agentic inequality" is not merely a capability gap; it is a structural bias in economic negotiations and transactions. When one party to a commercial interaction is represented by a sophisticated agent fleet and the other is operating manually, the information asymmetry and execution speed advantage of the agentic party is substantial.
"Agentic inequality is widening the gap between well-resourced entities that can deploy reliable, integrated agents and smaller players trapped by high-friction, low-trust tools — a structural bias that mirrors the early internet's winner-take-most dynamics."
This dynamic mirrors the early internet's winner-take-most economics, where first-mover advantages in network effects and data accumulation produced durable market concentration. The agent economy may produce similar concentration effects, but at the level of task execution rather than platform ownership.
Identity and Trust Infrastructure
Because agents now represent a large and growing share of internet traffic — executing transactions, sending communications, and making decisions on behalf of human principals — traditional identity infrastructure is becoming obsolete. The World Economic Forum has called for "Know Your Agent" (KYA) frameworks analogous to the Know Your Customer protocols developed for financial services in the 1970s: mechanisms to verify agent identity, authorisation scope, and accountability chains.
Without such infrastructure, the agent economy is vulnerable to systematic exploitation. Projections from security researchers suggest that one in four enterprise security breaches could stem from AI agent manipulation by 2028 — a figure that reflects not the inherent insecurity of agents but the absence of adequate identity and authorisation infrastructure.
Agentic inequality is widening the gap between well-resourced entities that can deploy reliable, integrated agents and smaller players trapped by high-friction, low-trust tools — a structural bias that mirrors the early internet's winner-take-most dynamics.
Governance Concentration Risk
The potential for dominant platform operators or governments to revoke agent access — effectively disabling the economic activity of organisations that have become dependent on agentic infrastructure — represents a systemic governance risk. This is not a hypothetical concern; it is the logical extension of the platform dependency dynamics that have characterised the internet economy for two decades, applied to a layer of infrastructure that is more deeply embedded in operational workflows than any previous technology.
Market Quality Mechanisms: Trust as Infrastructure
As the agent economy matures, trust mechanisms are emerging as critical market infrastructure. Marketplaces are implementing observability and logging systems that provide detailed audit trails of tool calls, latency, and costs; automated benchmarks that test agents against standardised datasets before publication; and escrow and insurance mechanisms that hold payments until tasks are completed or provide coverage for agent errors.
These mechanisms are not merely commercial features; they are the foundations of a functioning market. Without reliable quality signals, buyers cannot distinguish high-quality agents from low-quality ones, and the market degrades toward a lemons equilibrium in which low-quality providers drive out high-quality ones. The investment in trust infrastructure is therefore not optional; it is a precondition for market function.
The most sophisticated marketplaces are also implementing update cadence requirements — agents updated monthly rank significantly higher than those left untouched for 90 days or more — reflecting the recognition that agent quality degrades as the underlying models and data environments evolve. This creates a maintenance burden that favours well-resourced providers and may accelerate the concentration dynamics described above.
Governance Imperatives for the Agent Economy
The agent economy requires governance frameworks that do not yet exist at adequate scale. Several imperatives are clear from the current evidence:
Agent identity standards must be developed and adopted at industry scale. The absence of standardised mechanisms for verifying agent identity, authorisation scope, and accountability chains is the single most significant structural vulnerability in the current agent economy. This is a coordination problem that individual organisations cannot solve unilaterally; it requires industry-wide or regulatory action.
Outcome quality standards must be developed for high-stakes domains. When agents make consequential decisions in healthcare, finance, legal services, or public administration, the absence of outcome quality standards creates liability ambiguity and erodes public trust. Regulatory frameworks that specify minimum quality thresholds for agents operating in high-risk domains are a necessary complement to market-based quality mechanisms.
Agentic inequality must be addressed through structural interventions. If the agent economy produces the same winner-take-most dynamics as the platform economy, the distributional consequences will be severe. Policy interventions — including interoperability requirements, access mandates for small and medium enterprises, and antitrust scrutiny of agent marketplace concentration — may be necessary to prevent the agent economy from reproducing and amplifying existing economic inequalities.
Labour transition support must be scaled to match the pace of displacement. The junior crisis in knowledge work is accelerating faster than reskilling infrastructure can respond. Wage insurance, retention tax credits, and long-term income support mechanisms are being discussed in policy circles; the urgency of implementation is increasing.
Conclusion: The Architecture of the Agent Economy
The agent economy is not a future scenario; it is a present reality whose structural features are becoming legible. The shift from access-based to outcome-based pricing is restructuring how value is created and captured in digital markets. The emergence of three-sided agent marketplaces is creating new distribution dynamics with significant implications for market concentration. The empirical evidence of autonomous economic behaviour among agents is raising foundational questions about market design and governance. And the uneven distribution of agentic capability is creating structural inequalities that will compound over time without deliberate intervention.
The organisations and institutions that understand these structural dynamics — and act on them — will shape the agent economy's architecture. Those that treat it as merely a faster version of existing software automation will find themselves operating in a market whose rules have changed around them.
The agent economy is not automating the existing economy. It is building a new one on top of it. The question is not whether to participate, but on what terms.
