Author: Robbie Petersen , Junior Partner at Dragonfly

Compiled by: Gu Yu, ChainCatcher

Whenever a new narrative enters public discourse, the mainstream arguments are simplified into the form most easily accepted by the masses. Intuitively, when no one can empirically prove what will happen, provocation is more likely to be rewarded than meticulous analysis.

The recent discussions surrounding "smart agent commerce" are no exception. A consensus has emerged in the market: the number of smart agents is surging; smart agents need to conduct transactions; smart agents cannot hold bank accounts, but can hold e-wallets; card organizations charge 2-3% transaction fees; therefore, stablecoins are the winner.

This logical chain has flaws at many levels. Intelligent agents can hold bank accounts under an FBO (Financial Business Operator) architecture. Furthermore, the 2-3% transaction fee reflects credit and fraud risks, which blockchain cannot address.

However, the debate about "which payment method wins?" actually stems from a fundamental question that has been largely ignored in the discussion:

Do most intelligent agents actually engage in transactions?

The economic scale of intelligent agents will be enormous, but the proportion of intelligent agents actually engaging in transactions will not be that high.

The intelligent agent economy is more like an organizational chart than a market.

Fundamentally, artificial intelligence is a technology of automation. It can perform certain tasks—such as searching, aggregating, and synthesizing—and more efficiently than humans. Intelligent agents are the operational derivatives of artificial intelligence. They do not simply return outputs, but rather perform actual actions.

The underlying assumption of the entire intelligent agent business theory is that execution comes at a cost. In other words, for most intelligent agent tasks, they need to spend money to independently acquire external resources, pay for computing and data on a usage basis, and interact with other intelligent agents as independent economic agents.

This is fundamentally contradictory to the actual application of intelligent agents.

In general, the deployment of intelligent agents can be divided into two categories: business intelligent agents representing enterprise deployments, and consumer intelligent agents that enhance our personal lives. For different reasons, neither type of intelligent agent is likely to engage in autonomous transactions.

Commercial agency is an inevitable evolution of SaaS.

A sound business agent concept is an inevitable evolution of SaaS. They don't enhance workflows, but rather replace them. Just as over 95% of software spending comes from enterprises and governments, over 95% of large-scale agent applications are likely deployed within similar organizations.

This is the first subtle difference overlooked by current mainstream agency business theory: the vast majority of agency demand is not about agents booking tickets for consumers, but rather top-down deployments within enterprises. More importantly, there is a fundamental difference between an agent automating tasks within a closed organization and an agent operating as an independent economic entity.

Take a sales agent as an example. It integrates with the CRM system, researches potential customers, writes personalized marketing copy, and arranges follow-ups. It doesn't make any independent expenditures or interact with external agents from other organizations. It simply performs one task—sales development—within a closed environment and automates it.

Intuitively, this applies to almost all organizational functions. Financial agents review and reconcile expenses; accounting agents record journal entries, reconcile accounts, and prepare reports; legal agents review contracts and identify exceptions; coding agents write codes.

In almost all use cases, the agent itself does not spend, nor is it granted spending permissions. It is deployed top-down within a controlled organizational environment with access control mechanisms in place.

Even if it does require cross-organizational interaction and payment for its API calls or data, the cost may not be reflected in the form of agent-initiated payments. Any cost per use can be abstracted by the software vendor. This is exactly how enterprise software stacks operate. Platform providers negotiate customized partnerships with data vendors, computing providers, and other infrastructure partners, packaging access permissions into the platform cost and passing it out as a single, aggregated entry.

Furthermore, they achieve this at unit economics that no single agent can replicate autonomously. Computing resources are acquired through reserved capacity agreements with AWS, Azure, or GCP. Pricing for model inference is based on volume agreements with companies like Anthropic, OpenAI, or Google. Data augmentation is handled by vendors such as Bombora or Clearbit. All of this is pre-negotiated and abstracted.

In other words, the 40,000 API calls, model inferences, and data queries made by the agent will not generate 40,000 payments, but rather a single invoice. The granularity of consumption is never the same as the granularity of settlement, and businesses may prefer to maintain this state.

The consumer agent will be responsible for coordination, not consumption.

While commercial agents may not trade autonomously because businesses won't allow it, consumer agents also won't trade autonomously because people don't want them to.

To give an example that smart business advocates like to cite: You ask your agent to book a trip to Tokyo. It searches hundreds of hotels, cross-references reviews, checks your calendar, and applies your preferences. Then, it automatically books the room. You don't have to do anything. Of course, those who tout the agent-based business model will extend this user experience to almost every consumer sector, from groceries to home goods to clothing, and so on.

The problem is that preferences are not static. They are reflected in the act of making choices themselves. When you book a hotel, you're not just looking for the lowest price. Your judgment reflects your mood, the situation, your risk tolerance, and other qualitative factors that you might not have been aware of before looking at the options.

In practice, the agent will search, ask follow-up questions, and return options. You'll look at images, inquire about the surrounding environment, and perhaps read some comments. Then you'll make a choice and authorize the agent to use the credit card information they have to make the payment. In other words, the agent is merely a research assistant, not an independent economic entity.

Aside from certain predictable repeat purchases, this user experience is likely to remain consistent across almost all consumer sectors, precisely because consumer decisions are rarely based solely on price. The entire consumer economy is built on product differentiation. Whether it's clothing, hotels, home furnishings, or groceries, the decision-making process involves countless qualitative factors that not only cannot be captured by intelligent agents—but more importantly, these factors are inherent in the user's discovery process.

Intelligent agents will play a powerful coordinating role during the discovery phase, but at crucial moments, they will relinquish decision-making power to humans. Semantically speaking, this is not a business of intelligent agents, nor does it require the establishment of new payment channels.

The real winner of crypto payments: bottom-up agents

Although these two types of agents may account for more than 95% of agent deployments in the next five years, there is a third type worth paying attention to.

In recent months, a new type of bottom-up agent has emerged. Driven by the OpenClaw phenomenon, these agents belong to a distinctly different category. Unlike the aforementioned commercial and consumer agents, they are truly autonomous actors, independent of any organizational entity. These agents require actual payments, and the granularity and frequency of these payments make manual authorization impractical. Although the bottom-up agent economy is currently very small, it is likely to grow rapidly with the emergence of some unexpected new use cases.

Therefore, the debate about whether crypto payments or card networks are the best underlying architecture is only convincing within this extremely narrow context. While everyone is citing technical arguments for the superiority of crypto payments, it seems to me that their ultimate victory may be due to something else—permissionlessness.

Currently, the reality is that neither of these payment methods is technically optimized for agent-based commerce. While blockchain theoretically offers better unit economics for small payments, it lacks identity verification and risk scoring mechanisms—which may be particularly important in the future era of agents. Furthermore, while instant settlement is frequently mentioned, it merely means that fraudulent transactions are settled immediately on-chain. In contrast, card organizations possess complex fraud graphs and tokenized credentials that agents can inherit, but these tools are trained based on human behavioral patterns and cannot be directly mapped to autonomous agent transactions. Moreover, for cross-border transactions, agents are also subject to the settlement time constraints imposed by card organizations.

Perhaps contrary to intuition, the reason why encrypted payment methods may become the default infrastructure for such smart agents is because blockchain is open, permissionless, and unregulated.

This is its ultimate structural advantage. While I believe existing card organizations like Visa and Mastercard will continue to adapt through initiatives like Visa Intelligence Commerce and Mastercard's AgentPay, they are, after all, publicly traded companies and must comply with regulations, meet customer access requirements, and work with institutional counterparties. Blockchain, on the other hand, has no such limitations. Anyone can develop on a blockchain, any agent can transact, and no approval is required.

Intuition tells us that emerging, experimental categories will develop where there is the least friction.

The bottleneck lies not in infrastructure, but in ourselves.

However, a longer-term question is how this experimental pace of development can ultimately have a greater impact. Bottom-up agent economics will only truly flourish when autonomous agent organizations are significantly superior to human organizations enhanced by agents; this advantage must be not weak, but significant enough that top-down human constraints on agents become a competitive disadvantage. At that point, agents will no longer be merely automated executors of human tasks in closed environments, but will become the organization itself.

However, we may still be far from that future. The bottleneck won't be the technology itself. Moreover, what's truly "unsuitable for machines" may not be the payment system itself, but everything else not designed for an autonomous agent economy: regulatory frameworks, institutional bureaucracy, legal structures, and the social inertia surrounding human decision-making. These constraints have far more profound implications than any technical detail in the payment stack. Unfortunately, protocol upgrades cannot solve these problems.

The agency economy will be very large, with most of it being billed on a monthly basis.