Đối thoại a16z Crypto: Thời đại AI mua sắm giúp bạn sẽ trở thành như thế nào?

Original video title: The end of ads? AI agents are about to change how we buy
Original video source: a16z crypto
Original compilation: Deep潮 TechFlow

Editor’s introduction

This episode features a16z Crypto CTO Eddy Lazzarin, partner Noah Levine, and former a16z colleague now founder of Agent Cash, Sam Ragsdale. The three discuss the current state of AI agents, payment infrastructure, and the fate of credit card systems in a dense, high-frequency conversation.

The core conclusion is that the real-time settlement and zero marginal cost features of stablecoins are naturally suited for microtransactions at the 1-2 cent level in an agent economy, while the transaction fee system of credit cards (2-3% marginal fee + 30 cents fixed fee) is no match in this world.

Agent Commerce is dismantling the advertising business model that has dominated the internet for 20 years. Eddy Lazzarin even boldly states: “The advertising economy contract is dead; it will be completely gone within 10 years.”

Highlights and quotes

The essence of AI agents

· “LLMs are chatbots; agents are chatbots that can operate your computer. What humans can do with computers, agents can do too.”

· “Since around November last year, AI models have become smarter. They can complete complex tasks over sufficiently long time spans and will use tools. We started calling them ‘agents’ because they do more than just write code—they help you complete entire tasks.”

· “Internally, we call this ‘instant natural language programming.’ Users describe their needs in natural language, and the agent writes a potentially thousand-line JavaScript program in the background to execute it, costing only 20 cents per token generation and 10 cents per API call, then discards the program after use. Four years ago, this would have required an expensive software engineer a week to complete.”

No front-end merchants and business restructuring

· “What does a ‘headless merchant’ look like? It’s oriented toward AI services, not humans. No website front-end, only API endpoints and well-written documentation so models can read, understand, and call.”

· “The leading data industry players charge at least 100 times less than their competitors, using the same downstream data sources. Their core product is actually enterprise sales teams, not the data itself. In a decision-making world driven by agents, agents aren’t fooled by attractive sales teams. They try all data sources, find the best and most cost-effective, and remember it.”

· “You excitedly let your agent run all night. When you wake at 9 am, it’s been stuck since 2:30 am because the next step requires you to call the enterprise sales team.”

The end of advertising models

· “The economic contract of the internet since 2000 has been based on distraction-based monetization. Agents don’t get distracted. If they visit your site for recipes, they won’t see shoe ads next to it. The old model will die within 10 years.”

· “In 2016, total internet advertising was about $60 billion, and everyone thought it had peaked. Today, Google alone earns $300 billion annually from ads. But after GPT-4 emerged, traffic on tech news sites dropped by about 80%, and Stack Overflow experienced the same. These early adopters have decided to use agents for information retrieval and code execution. Others will follow because the experience is genuinely better.”

Stablecoins vs Credit Cards

· “Average transaction size on Agent Cash is 1-2 cents. The fixed fee for credit cards is 30 cents. Transaction fees in this scenario are completely absurd. It’s 2026; loyalty should belong to merchants, not the card used for payment.”

· “Credit cards appeared earlier than the internet and successfully transitioned from non-internet to internet era. Despite being heavily disrupted, they survived. So the verdict is still open.”

· “If someone from a credit card company is listening, and you have a money transfer license, you could instantly mint stablecoins for clients, allowing them to pay with stablecoins. I strongly suggest you consider this.”

The future of consumer experience

· “If an agent is shopping for you, and you equip it with a credit card skill, you can now see the ROI of each card precisely. When you have zero loyalty to any card, all psychological lock-in effects disappear.”

· “One day, you’ll realize you never really liked shopping.”

Open architecture for agent commerce stack

Host: Hello everyone, today I’m joined by a16z Crypto CTO Eddy Lazzarin, partner Noah Levine, and former a16z Crypto colleague now founder of Merit Systems, Sam Ragsdale. He’s working on a project called Agent Cash, and we’ll dive into it shortly.

Before that, I want to set the background. The AI agent space is evolving so rapidly that unless you watch 24/7, you can’t keep up. So, what’s the current state? Sam, you’re on the front lines—can you start?

Sam Ragsdale: I like to start with a classification framework borrowed from Erik Reppel, co-creator of Coinbase x402 protocol.

This classification divides agent commerce into two types. The first is conversational commerce—checking out within ChatGPT. You tell ChatGPT: “I’m a man living in West Village, NYC, going to Equinox gym, want to buy a pair of shoes to fit into my social circle.” It will empathetically recommend Nike shoes, and you buy.

The second is entrusting money to an agent to spend on your behalf to complete tasks.

Conversational commerce will definitely happen. All cutting-edge models like ChatGPT, Gemini, Claude will have checkout functions. It benefits consumers by helping them find better options; merchants by increasing conversion; platforms by taking 5-10% cut. It’s like a new Google Shopping.

The other world is that agents’ capabilities are still limited. Many people ask agents to do difficult things, like “help me with sales outreach,” and the agent says “I can’t, I don’t have access to that info.” If an agent has a small balance and can spend a few cents to buy services it couldn’t access before, it becomes more powerful.

So, two worlds are running in parallel: one where traditional LLM interfaces recommend products and handle the last mile, with platforms taking a cut; and another where you deploy your own agent to buy goods and services for you.

Noah Levine: I see two versions. One is the natural evolution of e-commerce—platforms shifting, mobile commerce emerging, new ad formats and Google Shopping. People always buy; consumer behavior changes. Now, with information access via LLMs, commerce naturally migrates into agents.

The other less “anthropomorphic” version is that the internet itself is changing. How people get information and execute actions is shifting with LLMs. The internet we built over the past 20 years might not be the internet of the future.

Using Google search, clicking on a sales-heavy webpage UI—those paths may no longer matter. Instead, a more agent-native internet will emerge, where agents pay directly for what they need, making humans more efficient.

Host: That directly relates to your investment theme, Noah. But before expanding, I want to give listeners a basic primer. Everyone’s used to interacting with LLMs, but now we hear about tools like OpenAI’s Codex—these agents have significant autonomy and can really get things done. If you haven’t been paying close attention, you might not realize how far the technology has come. Eddy, can you explain?

Eddy Lazzarin: Let me quickly review the past five months. Starting around November-December last year, AI models became smarter. Specifically, they can complete complex tasks over long periods and will use tools. We started calling them “agents” as a personification—they do more than just code; they help you finish entire tasks.

But agents aren’t capable of everything. Software isn’t just a small program running on your computer. The internet tells us you need to connect many other things to do interesting stuff—networks, participants, services.

Agents solve intent-building problems, partly also modeling preferences. You tell it what you want, it understands your goal, maps it to tools, networks, services. Through dialogue and memory, it can roughly grasp your preferences and pass this intent to tools, software, and providers.

These two parts are solved, which is exciting. Everyone wants to solve the remaining issues, but they are complex. At least, if you want your agent to do transactions, you need to handle authorization and delegation: how to prove the agent represents you? How to manage identity and authentication?

Then there’s payment and settlement. Once connected, reflecting your intent and knowing what to do, it needs to pay, demonstrate payment capability, handle split payments, refunds, etc. I skipped search, anti-fraud, and other critical parts, but you see, once intent building and preference modeling are automated—tasks previously only humans could do—the entire commerce process can be automated. Engineers’ reactions: “Wow, these two things—input or at least verbalize—can now be automated. Amazing.”

When people talk about “agentic commerce,” they mean the chain from “I talk to the agent” to “it gets what I need” and the chain reaction, because many things will be fundamentally rewritten.

Host: Very helpful. So, we’ve evolved from natural language interaction with LLMs to an enhanced version that connects various networks and real-world systems.

Eddy Lazzarin: It’s not just about connection. It’s not that what changed is what it connects to. Your laptop was always connected to everything; the connection aspect hasn’t changed. What has changed is that now they can use tools, think long-term, and stubbornly keep trying until the task is done.

Sam Ragsdale: Let me simplify my simplified version. LLMs are chatbots, good at dialogue. In the past, people thought they were best for customer service. Once dialogue was perfected, we developed tool usage. To oversimplify, they learned to operate computers. LLMs are chatbots; agents are chatbots that can operate your computer.

The key is, around GPT-4, they reached human-average operational levels at a fraction of the cost—about 1,000 times cheaper—and can scale capabilities significantly with more money. So roughly, what humans can do with computers, agents can do too.

Eddy Lazzarin: Exactly. The premise is simple, but the changes it triggers are huge—short-term, medium-term, long-term. Short-term, everyone is working to connect pipelines so agents can actually do things. Long-term, if your agent can access apps, how many UI or interfaces do you need? Do you still need Amazon App? Maybe Amazon App can’t compete with having your agent do all the homework, read reviews, and only show you what matters.

Sam Ragsdale: We call this internally “Just-in-time Natural Language Programming,” though the name isn’t very catchy. It turns non-programmers into programmers. You say: “I want to buy something for my fiancée on Amazon, based on her preferences, which are similar to what I bought last time. Browse about 1,000 options, pick the best match, order, find my address, and ship.”

What actually happens is the agent writes a program internally to do this complex task—maybe a thousand lines of JavaScript and Bash. It executes, but the user doesn’t see it, then discards it after.

Four years ago, that was science fiction. Writing such a program would require an expensive software engineer a week to debug and get API keys. Now, it costs about 20 cents per token, plus 10 cents per API call, and the program is thrown away after use—so cheap that no GitHub upload is needed. Even non-technical people can do it. My parents are now writing natural language programs—they don’t even realize it. They might now call themselves software engineers.

Host: That’s crazy. Are you engaged? Was that example your real experience?

Sam Ragsdale: I am engaged, thank you. But the ring wasn’t bought by AI. That ring predates AI—probably even the first computer.

On “Headless Merchants”

Host: Great, now let’s talk about these chain reactions. Sam, you previously mentioned how business will change in a world where agents handle large volumes of transactions, which directly relates to your concept of “Headless Merchants.” Can you explain what a headless merchant is?

Sam Ragsdale: Sure. I think it’s necessary to step back a bit. Besides traditional shopping scenarios like buying shoes via ChatGPT, there’s a huge B2B developer tools market. Platforms like Claude Code, OpenAI Codex are democratizing everything—anyone with a computer and tokens can build.

In the past, experienced developers would choose tools with clear preferences, follow enterprise sales processes, sign contracts. Now, new developers come with only “what I want to do” intent, no fixed idea about resources. Their creations are highly temporary, pay-as-you-go services that don’t require months of onboarding.

So, what does a headless merchant look like? It’s oriented toward AI services, not humans. No physical or digital storefront for browsing; only an API endpoint and good documentation so models can read, understand, and call.

Eddy Lazzarin: I resonate deeply. I feel like I was probably an AI in a past life. As a software engineer, I always think: if I visit a website and see no pricing, no way to get an API key with a credit card, I close the page. I don’t want to talk to sales, send emails.

Scheduling meetings with sales is a huge commitment and slows things down. I want to try now, immediately. If I’m working on something over the weekend and want to publish Monday, I just swipe my credit card, get the key, reimburse later, plan for later. That’s fast.

In the era of instant and temporary software, do you really want your agent to wait? It runs all night, then at 9 am you see it’s been stuck since 2:30 am because the next step requires talking to a sales team.

Sam Ragsdale: Not to mention, if the onboarding involves enterprise sales, the API cost could be ten times higher, because they need to manage customer relations.

Eddy Lazzarin: Completely unacceptable. You want your agent to run autonomously—not because you don’t care about what it does, but because you need speed, testing, quick iteration based on user feedback. You can’t wait.

If an AI model sees three options: one requiring enterprise sales contact, one needing a dedicated credit card, and one simply sending some stablecoins to get $10 worth of tokens for proof-of-concept, it will always choose the third. That alone could trigger some market restructuring.

Host: For traditional companies, these frictions make business harder, but they also lock in customers and loyalty. If these frictions disappear, how can revenue be reliably predicted?

Eddy Lazzarin: Here’s my bold take: let’s break everything. Add friction everywhere, make things hard to use. What are we doing?

I say this because friction can sometimes be useful—like blocking spam, creating filtering effects. But friction also has huge costs. As the economy accelerates, productivity rises, and the leverage of every minute increases, the opportunity cost of friction is rising. That’s the trend now.

Back to the core: even in the lowest-friction environment, if you get an API key instantly—or even better, pay directly with crypto wallets, where the wallet address is your account—you’ll still find other things that make services sticky.

Reputation, memory, state, data, and even intangible trust in the agent. If the agent knows you need answers fast, wants to move quickly, it won’t spend 20 minutes exploring all new options. It will remember what worked well last time and reuse it—like a smart person.

Sam Ragsdale: Let me give a down-to-earth example. We communicate daily with many merchants, most of whom have tried all current API-based sales channels. We’ve talked to many about their approach to “agent-native distribution,” a native way to distribute to AI agents.

Data products are usually commodities, with 5 to 50 sellers. The top one earns the most, charging about 100 times less than competitors, often using the same downstream data sources. Their core is enterprise sales teams, not the data itself. In an agent-driven decision world, agents aren’t fooled by attractive salespeople. They try all sources, find the best and cheapest, and remember it.

They do this by relying on enterprise sales teams. These teams are usually well-dressed, fly to your office, demonstrate: “Look at our beautiful data, no one’s better, $35,000 a year.” You sign a two-year contract, then they come back at renewal, repeat the pitch. Thousands of companies pay this way.

Smaller, better products with easier-to-use data often fail because they lack distribution channels. No innovation here—enterprise sales are the core product, data is secondary.

In an agent-based selection world, agents won’t want to chat with salespeople or be fooled by slick pitches.

They will try all data sources, find the most effective and cost-efficient (especially in bulk), and store that in memory: “Next time I need this data, I’ll use Minerva, not the other three.” This creates a more efficient world. Companies that used to pay $35,000 can now spend that money elsewhere.

Noah Levine: Another perspective: if you believe AI will spawn many solo or tiny teams capable of building what previously required 50-100 people, then enterprise sales teams flying to a single person’s basement makes no sense.

On one hand, existing merchants worry about revenue predictability—correct, change causes resistance. But on the other hand, it’s a new customer acquisition funnel. Reducing onboarding friction and tools access creates huge opportunities.

Sam Ragsdale: Most users have never used APIs, don’t know what they are, never got an API key, never signed enterprise agreements. But their first time, they can combine six different merchants’ APIs, write a natural language program, complete a task, then discard it. This creates a new market for API consumers.

The existing internet business model will be reshaped

Host: Sounds like Clayton Christensen’s innovator’s dilemma—high-end market sells expensive software to big clients, low-end market is new users trying out agents. But what can turn these toys into something truly disruptive?

Sam Ragsdale: Because it will ultimately deliver a better experience.

Noah Levine: I’d add: although it looks experimental today, history shows similar platform shifts. Stripe started serving tiny, long-tail merchants, many of whom grew into giants—that’s why Stripe keeps growing.

Shopify too—initially dropshipping and T-shirts, now serving many brands that built large businesses on Shopify. Similarly, a new wave of lean developers will use AI to build big companies. Tools they buy in agent commerce today will become huge consumption as they grow.

Sam Ragsdale: That e-commerce perspective is good. But I want to go bigger: the internet’s economic contract is dead.

Since Google launched in 2000 as the biggest promoter of a “free, open internet,” the contract was: you publish good content, people find it, Google displays it.

Later, AdWords introduced banner ads. The contract shifted: you publish good content, users land on your site, you show small ads, Google shares revenue. You can publish anything people want to see; Google manages advertisers and gives you a cut.

In this process, Google became the biggest driver of a “free, open internet.” They want the internet fast, cheap, everywhere—because the more you search, the more they earn.

Fundamentally, the internet’s business model is “distraction.” When you consume content—search info, recipes, scores—you get distracted. Later, you might buy those shoes or learn about a new B2B SaaS.

This scale has grown beyond expectations. I checked the 2016 “Internet Trends Report”: total ad spend was $60 billion, and people said “it’s peaked.” But today, Google alone earns $300 billion annually from ads.

But after agents appeared, search, info retrieval, and execution are shifting to agents. It’s still early; ChatGPT has 100 million monthly active users, but they’re still using it like Google search. They haven’t yet adopted agent-style use—like “help my dad find a Father’s Day gift and order.”

This is on the way. Data from the tech sector: since GPT-4, traffic on tech news sites has dropped about 80%, Stack Overflow the same. These early adopters are already using agents for info and code. Others will follow because the experience is better.

Old business models are being abandoned. Agents don’t get distracted. If they visit your site for recipes, they won’t see shoe ads. Publishers gain nothing. A new contract will be needed—one that justifies serving agent requests, not ads.

Will it be paid articles? Not sure. Paid API resources? Not sure. Will the internet look completely different? Also uncertain. But the old model will die within 10 years.

Host: If the core of the internet’s business model is distraction, that’s interesting because Google originally opposed portals. Yahoo and AOL offered links and content, Google just a search box, quick info. The evolution you describe is turning it into a distraction machine.

Now, we say agents won’t distract, but why would their evolution differ from humans? Could mechanisms be designed to lure agents, make them linger longer?

Eddy Lazzarin: That’s a big and fascinating question. The core is: who do agents represent? I recently heard someone say, “I’ve started using Google Search again because the top AI answers are good enough.” In that scenario, the “agent” is working for Google, running in Google’s cloud, controlled by Google. Would that agent be distracted by Google? I think yes.

The key is: whose goal function is it optimizing? Whose interests does it serve? If it’s serving Google’s interests, then it’s “distracted” from your benefit.

I don’t see it as entirely pessimistic. Good ads are good content—this industry consensus has existed for years. Good ads are almost indistinguishable from the content you want.

But let me clarify: if an agent works for Google or anyone else, the entire chain of commerce it participates in will be defined by them, using their methods and infrastructure that favor their business.

If an agent works for you, in extreme cases, running on your own open-source machine, fine-tuning it, changing system prompts—you can give it anti-distract tools. Then, advertisers face an opponent who can see through their tricks. I’m exaggerating, but adversarial scenarios could emerge.

Sam Ragsdale: Yes, there are countless ways to reintroduce ads. At the model weight level—most radical—by choosing training data that says “Nike is the best shoe in the world.” Nike could pay a billion dollars a year, and whether in ChatGPT or enterprise APIs, whenever shoes are mentioned, it’s Nike.

At the tool invocation layer, in system context, or as overlays that don’t even enter chat—model providers are grappling with this. Recently, Anthropic ran an ad mocking ChatGPT for advertising during the Super Bowl; OpenAI withdrew the ad afterward.

But OpenAI’s response was perfectly reasonable: “ChatGPT’s free users in Texas alone outnumber all of Anthropic’s paying users.” That’s a different scale—they need to serve many users unwilling to pay with expensive frontier tech, and advertising is a rational approach.

The reason ads are such a brilliant business model for search is because consumers don’t pay. High friction—like using credit cards—exists between advertisers, Google, and publishers, not with the billions of active search users. Those users get value just by opening Google.

If you align incentives, separate ads, and make them as relevant as possible, you get a better experience. Now, core model companies are moving away from ads. ChatGPT isn’t running ads; Gemini isn’t launching ads. Google probably will—because they have the scale, and they’ve done it before. They’re the biggest ad player. Gemini will eventually have ads, and Google Shopping-like features will appear.

But they know they don’t have a monopoly yet. All companies are competing, with lots of private market subsidies burning money. They don’t want to be seen as “less empathetic” or “less user-focused” because of ads. So, for now, no one is running ads; they’re trying to stay neutral.

Noah Levine: I see another direction: as merchants improve their pricing and product data transparency, they could shift ad spend into dedicated discounts for agent shopping scenarios. If agents are buyers, ad budgets could turn into discount budgets.

Another question is: what will the discovery layer of agent commerce look like? Who will do discovery? How to differentiate merchants? My prediction: if ads weaken because agents become buyers, and attention is no longer scarce, merchants might try discounts or subtle advertising by adjusting descriptions to make it easier for agents to understand.

Eddy Lazzarin: There are many dimensions. Advertising is just one way to get conversions. If a system can achieve higher conversions without ads, it will do so. There are many other ways: recommendation networks, discounts, coupons, special channels, free tokens for startups. Customer acquisition methods are numerous; ads are the most obvious because they’re the most directly felt.

If you turn the personalization knob all the way, and your agent has to chat with me first, it will tell you: “Eddy hates ads.”

The role of stablecoins vs credit cards in agent payments

Host: Before we finish, I have two questions. First: how well can traditional payment rails adapt to agent commerce? Or do we need entirely new native payment systems, like stablecoins, which seem to be finding product-market fit?

Sam Ragsdale: My overall judgment is that for e-commerce or conversational checkout scenarios, credit cards work well. They have built-in consumer protections—if the shoes don’t arrive or are damaged, Visa adjudicates, and you get your money back. The risk is on the merchant. This is a good deal for new goods and services.

But stablecoins are very useful in another class of scenarios. The average transaction on Agent Cash is 1-2 cents. About 600,000 such transactions have been completed. Credit card fixed fee is 30 cents. Wire transfer is about a dollar. Marginal fee rate is 2-3%, mostly for transaction fees, which give cashback points. For e-commerce, you might like points or miles; the 3% fee is borne by merchants. But when transactions are just 1-2 cents, with scattered API calls, stablecoins with zero marginal cost and fixed fees below 1 cent are ideal.

Another key point: instant settlement. When buying goods and services online, settlement cycles are at month-end—whether invoiced wire or credit card, merchants are essentially extending credit to customers or agents. In the agent world, you usually don’t know who the agent is.

Specifically, anyone who has used Anthropic or ChatGPT API knows the tiered system: first spend $50, settle once; then $100, then $2,500. This system exists because they’re providing credit—they don’t know you, haven’t done KYB or credit checks, and don’t know if you’ll pay at month-end.

AWS, Nvidia GPUs—same. Month-end settlement is terrible for these scenarios; merchants bear all the risk. If the customer isn’t a real company with a signed enterprise agreement, but an agent, you don’t know who it is. You could generate billions of agents overnight, but can’t extend credit to them.

Some are working on agent credit solutions, but I think that’s the wrong approach. Instant settlement directly solves the problem. It’s like cash—if I have it, I give it to you, you have it. You provide goods and services, I can’t claw back the money. Fixed fees with instant settlement are better for tiny transactions and this kind of activity.

Noah Levine: One point to challenge: the minimum transaction fee and whether credit cards can participate in microtransactions depend on card networks’ pricing.

If they want to introduce new transaction types—say “micro transactions”—with no minimum fees and lower rates, they can do it.

The benefit is that many consumers hold credit cards, more than those familiar with stablecoins. So, they can keep using cards for payments, and settle on the backend with stablecoins. But that will take time. Until then, native wallets using stablecoins directly for protocol consumption makes sense.

Sam Ragsdale: I think it’s very unlikely that credit card companies will overturn their 80-year core business model. But I’d love to see it happen.

Eddy Lazzarin: I agree—there’s no strict technical barrier for credit cards. But the issue is more subtle, involving business models and consumer perceptions. Recently, I saw concepts like “agent credit cards,” essentially virtual cards. I like the virtual card features from my issuer—generate temporary card numbers, cancel if fraud occurs or subscriptions are difficult to cancel.

Sometimes, new platforms or methods succeed not because they’re technically superior, but because they’re tailored for new scenarios. Credit cards are very old. They survived the transition from non-internet to internet, despite disruptions. So, the verdict is still open.

Noah Levine: Also, if Apple Pay becomes a viable tech, it could enable agent commerce. Will this disrupt Visa or Mastercard? My intuition is that many B2B transactions are settled via bank wire between developers and enterprise APIs. If card networks can capture this volume, through micro-

Link to original article

Learn more about Rhythm BlockBeats job openings

Join the Rhythm BlockBeats official community:

Telegram subscription group: https://t.me/theblockbeats

Telegram discussion group: https://t.me/BlockBeats_App

Twitter official account: https://twitter.com/BlockBeatsAsia