Lincoln Marr asked an AI agent to send Twitter articles to his Kindle by copying a trick he saw suggested online. When Twitter started blocking articles, the agent used Firecrawl to scrape articles and upload the results to a service called Stable Upload, where they were delivered to his email.
Ma, who heads up the AI products work at Coinbase, told CryptoSlate that he really forgot that the agents even had cryptocurrency wallets.
Firecrawl and Stable Upload are both part of a growing list of services built around x402. x402 is an open payment standard that allows software to pay for access the moment it is needed.
His agent selected both providers, paid them piecemeal, and completed the task without getting a single approval from him along the way.
When an agent chooses a vendor
Today, most AI tools work because the developer opened an account, generated an API key, and funded the credits.
Ma explained that there is an established pattern in which agents given advanced tasks can purchase whatever capabilities they need during their work.
He gave a clear example of a user asking an agent to convert a PDF to a podcast. Agents weigh Google’s NotebookLM and Celebrities in terms of cost, quality, and response time, and choose the one that’s right for the task, even receiving first-time user discounts along the way.
Murr pointed out that the wallet acts as both an agent identifier and a payment method, a combination that replaces the entire ritual of agents signing up for API keys and pre-funding their accounts before touching a new service.
The Linux Foundation launched the x402 Foundation in April, and Coinbase contributed the original protocol with early backers including AWS, Stripe, Visa, Mastercard, American Express, Shopify, Google, Microsoft, and Cloudflare.
Coinbase’s Bazaar currently indexes more than 10,000 paid tools that agents can search for and call directly, functioning much like Ma describes. That is, a directory built with the agent as the primary user.
In June, AWS added the ability for CloudFront customers to return machine-readable prices and payment terms to bots requesting protected content, and to verify payment at the network edge before granting access.
Cloudflare released its version in July, clearly explaining the new configuration where agents become buyers and requests become transactions.
From commitment to practicality
Marr framed the underlying movement as a shift from an economy built on attention to one built on utility.
He said the Internet runs a “Napster, Limewire” version of agent access, where agents are free to scrape sites because there is no software equivalent to advertising.
Cloudflare and AWS together account for about half of the world’s internet traffic, and they charge agents a few cents for the privilege of reading a page, giving them the power to alter traffic.
Ma argued that a similar shift could fundamentally hollow out subscription models, as flat monthly fees primarily benefit companies that collect the fees. In a world where a million agents pay a penny per call, something akin to walk-in retail opens up: frictionless, one-time, and available to everyone.
He expects that once microtransactions serve as the top of the funnel and paid usage patterns build trust on both sides, agents will gradually move toward more stable vendor relationships, discounts, and bulk pricing.
Who manages the directory?
In an economy where agents choose which services to pay for, the question arises as to who ultimately decides what gets discovered and ranked.
Ma pointed to x402’s governance as an answer, a foundation built from companies like Stripe and Coinbase that would normally be competitors.
Cloudflare will continue to control what happens with the sites behind its infrastructure, and its broader goal is to build payments into the very plumbing of the internet as a way for HTTP to become a shared standard that no one owns.
x402 processed approximately 75 million payments totaling $24 million in the last 30 days, with an average payment amount of approximately $0.32 per payment.
However, a population size study published in July complicated the picture, finding that the majority of on-chain x402 payment activity was fictitious or occurred within linked internal clusters.
Truly independent economic activity lies somewhere in between.
This investigation identified only $187,861 in identifiable independent service payments. The remaining $20.07 million may be genuine, but researchers were unable to rule out undiscovered connections between the wallets involved.
Ma said Coinbase wants to be the backbone of what he calls an agent economy, holding accounts, payment rails, and a discovery layer all at once, noting that not enough people currently carry around wallets to make these small purchases on a daily basis.
The challenges ahead for agency spending
Agent trading is where we see this tested, as the trading agent’s decisions generate numbers that can be directly evaluated against the cost of the data purchased.
Ma referenced Ethereum’s Vitalik Buterin’s long-standing idea that AI agents will eventually become the primary traders in illiquid prediction markets because their research costs are so low that they are not worth a human’s time.
Coinbase is building a product that allows agents to trade and pay directly from their retail accounts, and Ma expects the resulting loop, where agents pay for premium data and use it to improve their trades and bring more data providers into orbit, will gain real traction within six months.
Budget-conscious routing becomes a real advantage when agents can confidently decide whether their next purchase is worth it. As we found in a recent paper comparing free and paid search, cheap and expensive models, raw and validated data, and reasoning about budget constraints, smarter allocation beats a brute-force approach that uses four times as many resources.
Its world’s most powerful AI product doesn’t just run the largest model available, it manages a portfolio of paid inputs.
What a paid input agent should quote When is it worth paying?Search by failure modeAre new sources likely to change the answer?If confidence is low, or if the topic moves quicklyA premium to pay repeatedly for duplicate information Data Does your own data improve decision-making enough to justify the cost? If the task has measurable financial or strategic value Don’t add an edge Buy “premium” data A more powerful model Does this step require better inference? If the task requires synthesis, judgment, or complexity If the cost of being wrong is high. Less investment in checks to save pennies. Professional software. Does a dedicated tool beat common reasoning? If the task requires scraping, storage, formatting, trading, or execution. Choosing a cheap but unreliable tool.
If wallets, trust, and reliable value estimates are still lacking, as the July financial results survey has already shown, most transactions will be limited to demos, internal testing, and narrow developer workflows, and paid access infrastructure will continue to outpace usage.
Ma’s admission of weak demand provides a more accurate reading of the current situation. In other words, rails exist long before the economy that runs on them takes off.
Every part of that system, the wallet, the discovery layer, the payment rail, exists to answer whether the next cent an agent spends will improve their results. Ma, along with others rushing to build this economy, is pursuing the real answers to that question, and the answers are still taking shape.
