It took me a year to see the heartbreaking truth about Agent payments.

Author: Jessy; Translation: Jiahua, ChainCatcher

Over the past year, I have been dedicated to building infrastructure for the Agent economy, engaging with teams from startups driving Agent business such as Stripe, Visa, Coinbase, Google, and dozens of others. I have mapped out the entire industry, launched products, and tried to find market fit.

Currently, there is no genuine demand, and startups face many structural issues when venturing into this field.

Last month, Stripe announced 288 new products at the Sessions conference, with nearly 40% of the total document views coming from their Agent documentation. Their Agent business market has over 1,000 active merchants. However, at Sessions, the number of registered Agents conducting transactions was only in the single digits.

Visa mentioned that their Agent payment tokens (tokens bound to Agents, used for tokenized payments on behalf of users) currently require 3 to 9 months of KYC approval, and in practice, one must reach a minimum revenue threshold of $250 million to qualify. Today, only companies at Amazon and Walmart levels can complete this identity verification loop.

Coinbase reported that as of April, there are 69k active Agents and 165 million transactions on the x402 protocol. But independent on-chain analysis shows the actual daily transaction volume is about $17k, with roughly half being test transactions (according to CoinDesk, March 2026 report).

Agent for Merchants

We built shop.fast.xyz to directly verify the real-world application of agency-style commerce. It includes real products, merchants, and transactions.

For most product categories, the current AI shopping user experience is far inferior to traditional e-commerce. When buying clothes, electronics, or furniture, you want to see pictures, browse various options, and compare horizontally.

Conversely, dialogue-based chatbot interactions are a step backward. You’re replacing rich visual interfaces with plain text conversations, while humans are fundamentally visual shoppers.

Agents perform well in areas we initially thought would be difficult. They can understand user needs and handle instructions like "something similar but cheaper." The model layer plays a role here.

But it cannot replace the experience of browsing ten products side-by-side and choosing one. Chat interfaces can be enhanced with carousels and interactive displays, but at that point, you’re essentially rebuilding an e-commerce frontend inside a chat window. For visually driven price comparison shopping, we haven’t found a convincing reason why chat interfaces are better than native e-commerce interfaces.

We see genuine needs from merchants, but these are defensive in nature.

Merchants want their stores to be queryable by Agents. Not because current customers are buying through Agents, but because they worry that if this becomes a mainstream channel, they will be left behind.

This is an "Agent Engine Optimization (AEO)" strategy, but currently just an embellishment, not essential. Merchants are preparing for a wave that has not yet arrived.

Conversational commerce can indeed improve experience in certain scenarios: high-frequency, low-decision-cost purchases where users already know what they want. The most obvious example is ordering takeout. The market is huge, with high frequency and rapid decisions ("Help me order the same pad Thai from that last restaurant"). Here, dialogue-based Agents have an advantage.

But major food delivery platforms do not open their APIs. The only way is "computer usage": letting AI operate applications visually like humans. This approach is slow, fragile, and for a $15 lunch order, the reasoning cost is simply too high.

Another breakthrough lies in: some stores have extremely complex UI navigation, making the experience painful—layers of discounts, promo codes, loyalty programs, and confusing checkout flows.

An Agent that understands "use my coupons, deduct my reward points, find the cheapest shipping, operate in my native language" could simplify these currently poor experiences. This is especially important for elderly users, non-native speakers shopping at remote stores, or very niche scenarios with specific needs.

Both breakthroughs require large-scale consumer (B2C) distribution channels. You are competing with DoorDash (the largest food delivery platform in the US, holding 56% market share) and Amazon for user entry points.

Mass consumer distribution is an advantage of giants. The supply side of agency-style commerce is ready; demand is limited by user experience and distribution channels. Building more infrastructure alone cannot solve these two issues.

Agent for API

We discussed actual payment needs with dozens of developers. The situation is almost astonishingly consistent: today, Agents frequently use APIs for computation, reasoning, and data sources. Developers already have subscription services, archived API keys, and billing relationships with core providers.

A typical argument for stablecoins is: on Stripe, the minimum effective cost for credit card processing is about 2.9% plus 30 cents, making API calls under one dollar uneconomical. But for current low-frequency transactions, prepaid credit limits solve this problem. Developers top up their accounts in advance, and the issue is resolved.

Deeper issues lie in the vendor market. Most mainstream SaaS companies do not want to offer temporary API access costing only a few cents. Their business models rely on multi-year enterprise contracts. Those whose revenue depends on large commitment contracts will resist pricing mechanisms that bypass their existing models.

Machine-based commerce is inherently a long-tail market, including smaller services, niche data sources, individual developers, and MCP servers. Protocols like MPP and x402 are very suitable for this segment.

But by definition, this is a market serving high-end users with special needs, and historically, developers are often among the least willing to pay groups.

When Stripe Projects launched, it partnered with 32 vendors like Vercel, Supabase, Cloudflare, Twilio, covering most tools used by developers to build and deploy software, all accessible through existing billing systems. The top-tier needs of developer tech stacks are already met.

Opportunities for new payment channels exist beyond these top 30 services: real opportunities, but their scale is fundamentally much smaller than what the impressive numbers suggest.

The same pattern applies to content acquisition. Agents are already crawling and summarizing articles, while publishers are fighting back.

But when content monetization reaches large scale, it will be realized through CDN providers that are already between publishers and the internet (Cloudflare has launched AI auditing tools for this), or through large-scale licensing agreements between publishers and AI labs.

The infrastructure opportunities will ultimately flow to giants who already have distribution channels.

Agent for Agent

The business model of Agent-to-Agent is a long-term vision, currently almost entirely theoretical, with no meaningful transaction volume achieved. Startups are tackling core challenges: discovery, trust-building, terms negotiation, and dispute resolution.

When this transaction structure truly materializes, it will be fundamentally different from existing payment rails. Neither party will involve human identities. Latency will be sub-second. Funds ranging from fractions of a cent to millions of dollars will flow within the same process.

There will also be multi-party settlement mechanisms, which do not fit the bilateral buy-sell model of current payment systems. Once this happens, we believe it will come quickly and at scale.

This is a long-term bet on dedicated settlement infrastructure, and it does exist. But "real long-term bets" and "current markets" are two different things.

For months, we have been advocating for this market and have built a complete infrastructure around it over the past few years. With our distributed network, it could theoretically scale to over 1 billion TPS, with latency under 50 milliseconds, and average consensus time of 10 milliseconds. But we must align with the current market reality.

Agent for Finance

This is arguably the only category with existing demand. The customer base already exists and has willingness to pay. Today, fund managers, finance teams, and DeFi users are paying for financial tools. Embedding AI into existing workflows is a natural product evolution.

Agent finance also creates entirely new behavior patterns. An Agent capable of real-time autonomous monitoring and rebalancing hundreds of positions operates in ways impossible for humans to manually replicate. It’s not just automation; it’s a substantial capability upgrade.

The challenge lies in the competitive landscape. The financial industry is heavily regulated and highly dependent on existing relationships. Established institutions have licenses, compliance infrastructure, and client relationships. Startups can find footholds in lightly regulated areas (like DeFi), in slow-moving giants, or in fields where AI can create capabilities that incumbents lack.

Compared to the other three categories, the competitive dynamics here favor mature companies, because layering AI on top of existing products and customer bases is much easier than starting from scratch.

The True Battleground

So, why are people still building these things? There are two reasons.

First is motivation. Industry giants have ample cash flow to bet on a future that may take years to materialize. For them, the cost of entering five years early is negligible, while missing a year could be devastating. So, they must build.

Second is cognitive blind spots. When your core business is payments, every problem looks like a payments problem. If the Agent economy needs a payment layer, just build that payment layer.

But payments are only one part of a much larger problem. The real challenge is coordinating work between Agents and humans, verifying work results, and settling outcomes. Payments are just one aspect of settlement. Settlement is just one aspect of collaboration. And collaboration is the real big cake.

Large-scale collaboration will naturally generate settlement mechanisms as a necessity. Payments are just one instrument in this symphony, not the entire movement. Companies solving collaboration problems will acquire payment businesses, not the other way around.

Most established companies are engaging in defensive development to prepare for a future of large-scale machine transactions. Since their funding runway is unlimited, timelines are less critical for them.

But startups do not have this luxury. We must seek the true market opportunities and cannot wait idly for the wave to hit.

A year of building has led us to an unexpected direction. The market activity there is real, growing rapidly, and still underserved. It exists outside the four categories we have outlined.