Author: Robbie Petersen, Junior Partner at Dragonfly
Compiled by: Guyu, ChainCatcher
Whenever an emerging narrative enters public discussion, the mainstream argument tends to be simplified into a form most easily accepted by the general public. Intuitively, when no one can empirically prove what will happen, provocation is more rewarding than nuanced analysis.
The recent discussion around the "Agent Economy" is no exception. A market consensus has formed: the number of agents will surge; agents need to conduct transactions; agents cannot hold bank accounts but can hold crypto wallets; card networks charge 2-3% fees; therefore, stablecoins win.
This chain of logic is flawed on many levels. Agents can hold bank accounts under an FBO (Financial Business Operator) structure. Furthermore, the 2-3% fee reflects credit risk and fraud risk, which blockchain does not solve.
However, the debate over "which payment method wins?" actually stems from a preliminary question that is largely overlooked in the discussion:
Will most agents actually conduct transactions?
The agent economy will be massive, but the proportion of agents that actually transact will not be.
The Agent Economy Resembles an Org Chart More Than a Market
Fundamentally, AI is an automation technology. It can perform certain tasks—such as searching, aggregating, and synthesizing—more efficiently than humans. Agents are the actionable derivatives of AI. They don't simply return outputs; they perform actual tasks.
The implicit assumption in the entire agent commerce theory is that execution comes at a cost. In other words, for most agent tasks, they need to spend money to autonomously acquire external resources, pay for compute and data on a usage basis, and interact with other agents as independent economic actors.
This is fundamentally at odds with how agents will actually be used.
Broadly, agent deployment falls into two categories: commercial agents deployed on behalf of businesses, and consumer agents that augment our personal lives. For different reasons, neither category is likely to transact autonomously.
Commercial Agents Are the Inevitable Evolution of SaaS
A plausible concept of a commercial agent is the inevitable evolution of SaaS. Rather than augmenting workflows, they replace existing workflows. Just as over 95% of software spending comes from businesses and governments, over 95% of large-scale agent applications will likely be deployed within similar organizations.
This is the first nuance the current mainstream agent commerce theory misses: the vast majority of agent demand won't be agents booking flights for consumers, but top-down deployments within businesses. More importantly, agents automating tasks within a closed organization are fundamentally different from agents operating as independent economic actors.
Take a sales agent. It plugs into the CRM, researches leads, writes personalized outreach, and schedules follow-ups. It doesn't spend autonomously, nor does it interact with external agents from other organizations. It simply performs a task—sales outreach—and automates it within a closed environment.
Intuitively, this applies to almost every organizational function. Finance agents review and reconcile expenses; accounting agents record journal entries, reconcile accounts, and prepare statements; legal agents review contracts and flag exceptions; coding agents write code.
In almost every use case, the agent itself does not spend, nor is it granted spending authority. It is deployed top-down in a controlled organizational environment with permission controls.
Even if it does require cross-organizational interaction and pays for its API calls or data, the cost likely won't manifest as the agent paying autonomously. Any pay-per-use costs would likely be abstracted away by the software vendor. This is precisely how the enterprise software stack works. Platform providers negotiate custom partnerships with data vendors, compute providers, and other infrastructure partners, bundle access into the platform cost, and pass it through as a single aggregated line item.
Furthermore, they can achieve unit economics that no single agent could replicate autonomously. Compute is acquired through reserved capacity agreements with AWS, Azure, or GCP. Model inference is priced based on volume agreements with Anthropic, OpenAI, or Google. Data enrichment is done through vendors like Bombora or Clearbit. All of this is pre-negotiated and abstracted.
In other words, an agent's 40,000 API calls, model inferences, and data queries don't generate 40,000 payments; they generate an invoice. The granularity of consumption has never been the same as the granularity of settlement, and businesses may prefer to keep it that way.
Consumer Agents Will Coordinate, Not Consume
While commercial agents might not transact autonomously because businesses won't allow it, consumer agents won't transact autonomously because people don't want them to.
Take an example proponents of agent commerce love to cite: you tell your agent to book a trip to Tokyo. It searches hundreds of hotels, cross-references reviews, checks your calendar, applies your preferences. Then, it books the room automatically. You do nothing. Proponents of the agent commerce model would extrapolate this user experience to almost every consumption category, from groceries to household goods to clothing and beyond.
The problem is that preferences aren't static. They are revealed in the act of choice itself. When you book a hotel, you're not just looking for the cheapest accommodation. The judgments you make reflect your mood, context, risk tolerance, and other qualitative factors you might not even be aware of before seeing the options.
In practice, the agent will search, ask follow-up questions, and return options. You'll look at pictures, ask about the neighborhood, maybe read a few reviews. Then you'll make a choice and authorize the agent to pay using the credit card information it already has. In other words, the agent is a research assistant, not an independent economic actor.
Beyond certain predictable, repeat purchases, this user experience will likely hold true for almost every consumption category precisely because consumer decisions are rarely based on price alone. The entire consumer economy is built on product differentiation. Whether it's clothing, hotels, household goods, or groceries, the decision-making process involves countless qualitative factors that are not only impossible for an agent to capture—more importantly, they are discovered by the user during the process itself.
Agents will play a powerful coordinating role in the discovery phase, but at the moment of truth, they will cede decision-making back to the human. Semantically, this isn't agent commerce, and it doesn't require building new payment rails.
Where Crypto Payments Truly Win: Bottom-Up Agents
While these two categories combined will likely account for over 95% of agent deployment in the next five years, there is a third category worth considering.
Over the past few months, a new type of bottom-up agent has begun to emerge. Driven by phenomena like OpenClaw, these agents belong to a qualitatively different category. Unlike the commercial and consumer agents described earlier, these are truly autonomous actors operating independently of any organizational principal. These agents need to actually pay, and they need to do so with a granularity and frequency that makes manual authorization impractical. Although the bottom-up agent economy is currently极小, it is likely to grow rapidly as new, unforeseen use cases emerge.
It is only in this extremely narrow context that the debate over whether crypto payments or card networks are the better underlying infrastructure becomes compelling. While the technical arguments for why crypto payments are superior are being enumerated, the reason they might ultimately win, in my view, is something else—permissionlessness.
Today, the reality is that neither payment method is technically optimized for agent commerce. While blockchains theoretically offer better unit economics for micropayments, they lack authentication and risk scoring—which may prove important in the future agent era. Also, while instant settlement is often cited, it simply means fraudulent transactions settle immediately on-chain. Conversely, card networks have sophisticated fraud graphs and tokenized credentials that agents could inherit, but these tools are trained on human behavior patterns and don't map cleanly to autonomous agent transactions. Furthermore, for cross-border transactions, agents would be subject to the settlement times of card networks.
Perhaps counterintuitively, crypto payment rails might become the default infrastructure for this class of agents because blockchains are open, permissionless, and unregulated.
This is their ultimate structural advantage. While I believe incumbent card networks like Visa and Mastercard will continue to adapt through initiatives like Visa Intelligence Commerce and Mastercard's AgentPay, they are, after all, public companies that must comply with regulatory obligations, meet customer onboarding requirements, and work with institutional counterparties. Blockchains have none of these constraints. Anyone can build on them, any agent can transact, and no approval is required.
Intuitively,新兴的, experimental categories develop where friction is lowest.
The Bottleneck Isn't the Infrastructure, It's Us
The longer-term question, however, is how fast this experimental development can ultimately scale to matter. The bottom-up agent economy only truly flourishes when autonomous agent organizations are demonstrably better than human organizations augmented by agents; not marginally better, but sufficiently better that the top-down human constraints on agents become a competitive disadvantage. At that point, agents cease to be mere automators of human tasks in closed environments and become the organization itself.
However, we are likely far from this future. The bottleneck won't be the technology itself. And what is truly "not built for machines" might not be the payment rails themselves, but everything else not designed for an autonomous agent economy: regulatory frameworks, institutional bureaucracy, legal structures, and the social inertia around human decision-making. These constraints are far more profound than any technical detail in the payment stack. Unfortunately, protocol upgrades can't solve for them.
The agent economy will be massive, and most of it will be billed monthly.
