Thanks to @austingriffith, @dongossen, @ekrahm, @programmer, @RosuGrigore, @0xfishylosopher and @JackSimison for the review and feedback.
I've spent the last year building infrastructure for the agent economy, talking to the teams shipping agentic commerce at Stripe, Visa, Coinbase, Google, and dozens of startups. I mapped the space, launched products, and tried to find the market.
No real demand exists yet and there are many structural problems for startups to tackle this space.
Stripe shipped 288 new products at Sessions last month with agent doc traffic approaching 40% of documentation reads. Their agentic commerce marketplace has over 1,000 enabled merchants. At Sessions, the number of registered agents transacting was in the single digits.
Visa mentioned that their agentic token currently requires 3-to-9 months of KYC approval and effectively a $250M revenue minimum to qualify. Today, only Amazon and Walmart-class companies can close the identity loop.
Coinbase reported 69,000 active agents and 165 million transactions on x402 by April. Independent onchain analysis shows real daily volume around $17,000, roughly half of which is test transactions (CoinDesk, March 2026).
what we learned building shop.fast.xyz (agent-to-merchant, aka proxy commerce)
We built shop.fast.xyz - https://x.com/fastxyz/status/2061863041562755403 - to confront proxy commerce directly. Real products, merchants, and transactions.
However for most product categories, the current AI shopping UX is strictly worse than traditional e-commerce. When you shop for clothes, electronics, or furniture, you want to see images, browse options, compare side by side. A chatbot conversation is a downgrade. You're replacing a rich visual interface with a text thread. Humans shop with their eyes.
The agent was good at the part we expected to be hard. It understood what users wanted. It handled "something like this but cheaper" gracefully. The model layer worked. What it couldn't do was replace the experience of looking at ten products side by side and picking one. Chat can be enhanced with carousels and interactive displays, but at that point you're rebuilding the e-commerce frontend inside a chat window. For visual comparison shopping, we haven't found a convincing answer for why the chat wrapper is better than the original.
We saw real demand from merchants, but it was defensive. Merchants want storefronts to be agent-queryable. Not because customers are buying through agents today, but because they're worried about being left behind if it becomes the dominant channel. This is the Agentic Engine Optimization play, but it's currently a nice-to-have, not must-have. Merchants preparing for a wave that hasn't arrived.
Where conversational commerce does improve the experience: high-frequency, low-consideration purchases where the user already knows what they want. Food ordering is the clearest case. Large market, daily frequency, fast decisions ("order me pad thai from the place I liked last time"). A conversational agent could win here. But the major delivery platforms don't expose APIs. The only path is computer use - having AI visually navigate the app like a human. Slow, fragile, and the inference costs don't work for a $15 lunch order.
The other opening: stores complex enough that navigating the UI is genuinely painful. Layered discounts, promo codes, loyalty programs, confusing checkout flows. An agent that understands "apply my coupon, use my rewards points, find cheapest shipping, do it in my language" simplifies something that's actually broken today. This matters especially for older users and non-native speakers shopping on stores built for a different region, or under very specific scenarios where you're just having some really niche requests.
Both openings require massive B2C distribution. You're competing with DoorDash and Amazon for the user's entry point. Distribution at consumer scale is an incumbent advantage. The supply side of proxy commerce is ready. The demand side is constrained by UX and distribution, and more infrastructure won't fix either.
what we learned about x402 and MPP (agent-to-web/API, aka machine commerce)
We talked to dozens of developers about their actual payment needs. The pattern was almost universal: agent API usage today is recurring. Compute, inference, data feeds. Developers already have subscriptions, API keys on file, billing relationships with their core providers.
The typical argument for stablecoins: card processing has a minimum effective cost of roughly 2.9% plus 30¢ on Stripe, which makes sub-dollar API calls uneconomical. For today's low volume, topping up credits solves this. Developers pre-fund accounts and the problem goes away.
The deeper issue is the supplier market. Most major SaaS companies don't want ad hoc API access at fractions of a cent. Their business model is multi-year enterprise contracts. The companies whose revenue depends on large commitments will resist pricing models that route around them.
Machine commerce is structurally a long-tail market. Smaller services, niche data sources, individual developers, MCP servers. The protocols like MPP and x402 are well-designed for this segment. But it is by definition a market serving power users with specialized needs, and developers are historically among the least willing to pay.
Stripe Projects launched with 32 provider partners - Vercel, Supabase, Cloudflare, Twilio, and others - covering the majority of what developers use to build and deploy software, all accessible through existing billing. The top of the developer stack is already served. The opportunity for new rails is everything outside those top 30 services: real, but inherently smaller than the headline numbers suggest.
The same dynamic applies to content access. Agents already scrape and summarize articles constantly, and publishers are pushing back. But when content monetization arrives at scale, it will flow through CDN providers who already sit between publishers and the internet (Cloudflare has shipped AI audit tools for this) or through bulk licensing deals between publishers and AI labs. The infrastructure opportunity accrues to incumbents with existing distribution.
what we learned about agent-to-agent payments
Agent-to-agent commerce is the long-term vision and almost entirely theoretical. Nobody has shipped meaningful volume. The hard parts are being shipped right now by various startups: agent discovery, trust establishment, terms negotiation, dispute resolution.
The transaction structure, when it does materialize, looks nothing like existing rails. No human identity on either side. Sub-second latency. Values from fractions of a cent to millions in the same flow. Multi-party settlement that doesn't fit the bilateral buyer-seller model every existing rail assumes. When it does happen, we believe it’ll happen fast and in high magnitudes.
That's the long bet for dedicated settlement infrastructure, and it's real. But "real long bet" is different from "current market." We were one of the people claiming this market for several months and built a complete infrastructure around it the past years with our distributed network, which can theoretically scale up to more than 1B TPS with <50ms latency and 10ms on average consistency. But we need to meet where the market is now.
what we learned about agentic finance
Arguably the only category with existing demand. The customer already exists and already pays. Fund managers, treasury teams, and DeFi users all spend on financial tooling today. AI insertion into existing workflows is a natural product motion.
Agentic finance also creates entirely new behaviors. Agents that autonomously monitor and rebalance across hundreds of positions in real time operate in ways humans cannot replicate manually. There's actual capability uplift, not just automation.
The challenge is competitive dynamics. Finance is heavily regulated and depends on existing relationships. Incumbents have licenses, compliance infrastructure, and client relationships. Startups can carve out positions where regulation is lighter (DeFi), incumbents move slower, or AI creates capabilities incumbents don't have. But competitive dynamics here favor established players more than in the other three categories, as it's easier to add AI on top of existing product and customer base than the other way around.
the honest summary
So why is everyone still building this? Two reasons.
The first is incentives. The big players have the cash flow to bet on a future that takes years to manifest. Their cost of being five years too early is rounding error; their cost of being one year too late is catastrophic. They have to build.
The second is conceptual blinders. When your business is payments, every problem looks like a payment problem. The agent economy needs a payments layer; therefore, build the payments layer.
But payments are one piece of a bigger problem. The hard problem isn't moving money between agents. The hard problem is coordinating work between agents and humans, verifying what was done, and settling outcomes. Payments are one piece of settlement. Settlement is one piece of coordination. Coordination is the actual prize.
Coordination at scale naturally produces settlement mechanisms as a requirement. Payments emerge as one instrument in that orchestra rather than the entire composition. The companies that solve coordination will subsume payments, not the other way around.
Most incumbents are building defensively against a future where machines transact at scale, and the timeline doesn't matter to them because their runway is infinite.
Startups don't have that luxury. We have to find where the market is. We can't wait for the wave to break.
A year of building pointed us somewhere unexpected. The activity is real, growing fast, and underserved. It lives outside the four categories we mapped.
More on that soon.
