Agent Payment One-Year Observation: The Cold Reality Behind the Hot Narrative

Title: A Year Inside Agentic Payments: The Uncomfortable Truth

Author: Rhythm BlockBeats

Source:

Reprinted from: Mars Finance

Editor's Note: This article offers a relatively calm builder’s perspective: over the past year, intelligent agent payments have become a hot narrative at the intersection of AI, payments, and crypto. Companies like Stripe, Visa, Coinbase, Google, and dozens of startups are laying out plans. Concepts such as stablecoin micro-payments, x402, inter-machine settlement, and agent e-commerce are also gaining popularity. But after actually building products and engaging with merchants and developers, the author finds that real demand has not yet emerged at scale.

The article dissects several typical scenarios: agent shopping is not better than traditional e-commerce in most categories because users still need images, comparisons, and browsing; machine API payments seem suitable for stablecoin micro-payments, but most developers have already solved this through subscriptions, point recharge systems, and existing billing frameworks; payments between agents, while a long-term vision, are still in early stages with no real transaction volume.

In contrast, agent finance is one of the few areas with existing demand. Funds, treasury teams, and DeFi users already pay for financial tools, and AI can bring practical improvements like real-time monitoring and automated rebalancing. But this market also favors traditional institutions with licenses, compliance, and customer relationships.

The author’s final judgment is: the true missing piece in the agent economy isn’t just a simple payment layer, but more complex coordination capabilities—how to enable agents to collaborate with humans, verify task completion, and settle results. Payments are just one part of this. For giants, early deployment is a defensive move; for startups, the key is to find markets that already exist today.

Below is the original text:

Over the past year, I’ve been building infrastructure for the agent economy, and I’ve communicated with teams at Stripe, Visa, Coinbase, Google, and dozens of startups advancing agent-based commerce. I’ve mapped out this field, launched products, and tried to find the real market.

But the reality is: genuine demand has not yet appeared. For startups wanting to enter this space, many structural issues still exist.

Last month, Stripe announced 288 new products at the Sessions conference, and access to agent-related documentation has approached 40% of all documentation reads. Its agent business market has already onboarded over 1,000 merchants. But at the conference, the number of registered and transaction-completing agents was only in the single digits.

Visa mentioned that its agent token currently requires 3 to 9 months of KYC approval, and the basic requirement is that the enterprise’s annual revenue must be at least $250 million to qualify. Today, only companies at the level of Amazon or Walmart can close the identity verification loop.

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

What we learned from building shop.fast.xyz

Agent to merchant, or agent-based commerce

We built shop.fast.xyz to directly validate agent-based commerce: real products, real merchants, real transactions.

But for most product categories, current AI shopping experiences are clearly inferior to traditional e-commerce. When buying clothes, electronics, or furniture, users want to see images, browse options, and compare side-by-side. Chatbot-style conversations are actually a regression: replacing a rich visual interface with a string of text dialogues. Human shopping first and foremost is visual.

The agent performed well in what we initially thought was the hardest part. It can understand what users want and handle requests like “something like this but cheaper.” The model layer is effective. But it cannot replace the experience of “view ten products at once and pick one.” Chat interfaces can add product carousels and interactive displays, but at that point, you’re essentially rebuilding an e-commerce frontend inside a chat window. For shopping scenarios that require visual comparison, we haven’t yet found a convincing reason why a chat shell would be better than the original e-commerce interface.

We do see merchant-side demand, but it’s mostly defensive. Merchants want their stores to be queryable by agents—not because many consumers are shopping via agents today, but because they worry that if agent-based commerce becomes mainstream, they’ll be left behind. This is the so-called agentic engine optimization (Agent EO) opportunity, but it’s currently just “better,” not “necessary.” Merchants are preparing for a wave that hasn’t arrived yet.

The place where conversational commerce can truly improve experience is in high-frequency, low-decision-cost purchase scenarios where users already know what they want. The clearest example is ordering food. The market is large enough, the frequency high enough, and decisions quick enough—like “Order me the pad thai I liked last time.” In such scenarios, conversational agents might win. But major food delivery platforms don’t open APIs. The only path is computer use—making AI operate app interfaces visually like humans do. This process is slow, fragile, and for a $15 lunch, the reasoning cost simply doesn’t add up.

Another opportunity lies in complex online stores that cause user pain—layers of discounts, promo codes, loyalty points, chaotic checkout flows. An agent that understands “Help me use coupons, redeem points, find the cheapest shipping, and complete the operation in my language” could significantly simplify today’s broken shopping experience. This is especially important for elderly users, non-native speakers, or cross-region shopping, or in very niche scenarios with highly specific, complex needs.

But both opportunities require massive B2C distribution capabilities. You’re competing with DoorDash, Amazon, and others for user entry points. Consumer-scale distribution is an advantage for existing giants. The supply side of agent-based commerce is ready; demand is limited by user experience and distribution channels. More infrastructure alone can’t solve these two problems.

What we learned from x402 and MPP

Agent to Web/API, or machine commerce

We’ve spoken with dozens of developers about their real payment needs. The pattern is almost identical: today’s agent API usage is primarily for recurring consumption—compute power, inference, data sources. Developers already have subscriptions, API keys, account bindings, and billing relationships with core service providers.

A typical argument for stablecoin payments is that credit card payments on Stripe have a minimum effective cost of about 2.9% plus 30 cents, making API calls under $1 uneconomical. But at current low transaction volumes, recharging points solves this problem. Developers pre-fund their accounts, so this issue doesn’t arise.

Deeper issues lie in the vendor market. Most large SaaS companies don’t want to offer scattered API access at fractions of a cent. Their business models rely on multi-year enterprise contracts. Those dependent on large commitments will resist new pricing models that bypass this.

Machine commerce is structurally a long-tail market. It serves small services, vertical data sources, independent developers, MCP servers, etc. Protocols like MPP and x402 are very suitable for this niche. But by definition, it’s a market for specialized needs; developers have historically been among the least willing to pay.

When Stripe Projects launched, it integrated 32 service partners, including Vercel, Supabase, Cloudflare, Twilio, covering most core services used by developers to build and deploy software, all accessible via existing billing systems. The top of developers’ tech stacks is well served. The opportunity for new payment tracks lies beyond those 30+ service providers: it’s real but inherently smaller than the markets implied by grand narratives.

Content access follows the same logic. Agents are already crawling and summarizing articles, and publishers are fighting back. But when content monetization truly scales, it will likely happen through existing CDN providers between publishers and the internet—like Cloudflare’s AI audit tools—or via bulk licensing agreements between publishers and AI labs. Infrastructure opportunities will flow to those with existing distribution capabilities.

What we learned from agent-to-agent payments

Agent-to-agent commerce is a long-term vision, but currently almost entirely theoretical. No one has yet achieved meaningful transaction volume. The real challenge is being advanced by startups working on agent discovery, trust establishment, terms negotiation, and dispute resolution.

Once such transaction structures truly form, they will look very different from existing payment rails. Both parties will lack human identities; latency must be under a second; transaction amounts can range from fractions of a cent to millions of dollars; and multi-party settlement will be involved, not just the traditional bilateral buyer-seller model. When it finally happens, we believe it will explode rapidly and at a massive scale.

This is the long-term bet on dedicated settlement infrastructure, and it’s a real one. But “real long-term bet” and “current market” are not the same. We were among those claiming this market would arrive for months, and over the past few years, we built a whole infrastructure around it—including our distributed network. Theoretically, it can scale to over 1 billion TPS, with latency below 50 ms, and an average consensus time of 10 ms. But we must return to where the market is now.

What we learned from agent finance

This is arguably the only category with genuine existing demand. Customers already exist and are already paying. Fund managers, treasury teams, and DeFi users are already spending on financial tools today. Integrating AI into existing workflows is a natural product path.

Agent finance will also create entirely new behaviors. Agents capable of autonomous monitoring and real-time rebalancing of hundreds of positions can operate in ways impossible for humans to manually replicate. There is real productivity enhancement here, not just automation.

The challenge lies in the competitive landscape. Finance is highly regulated and depends on existing relationships. Existing institutions have licenses, compliance infrastructure, and customer bases. Startups can enter lightly regulated areas like DeFi or seek segments where incumbents are slow, or where AI can create capabilities giants don’t yet have. But overall, this field’s competitive dynamics favor established players more than the previous three categories, because adding AI on top of existing products and customer bases is easier than starting from scratch.

Honest summary

So why are people still doing this? Two reasons.

First is incentives. Large companies have enough cash flow to bet on a future that takes years to materialize. For them, the cost of entering five years early is just a tiny fraction; the cost of being a year late could be disastrous. So they must act.

Second is cognitive blind spots. When your core business is payments, every problem looks like a payments problem. Agent economy needs a payment layer, so everyone builds one.

But payments are just a part of a bigger problem. The real challenge isn’t just moving money between agents; it’s coordinating work between agents and humans, verifying task completion, and settling results. Payments are just one part of settlement. Settlement is just one part of coordination. And coordination is the real prize.

Large-scale coordination naturally creates demand for settlement mechanisms. Payments will become one instrument in this orchestration, not the entire composition. Companies that truly solve coordination will eventually incorporate payments into that, rather than being acquired by payment companies to handle coordination.

Most existing giants are defensively building a future of “machine-scale trading.” For them, timelines don’t matter much—they have nearly unlimited runway.

But startups don’t have that luxury. We must find where the market is right now. We can’t keep waiting for the wave.

A year of building has led us to an unexpected direction. There is activity, and it’s growing fast, but underserved. It exists outside the four categories we’ve mapped out.