AI agents can already hold money. What they have lacked is a web that lets them spend it. x402, an open payment standard that Coinbase launched with partner organizations, fills that gap by putting a long-unused part of HTTP, the protocol behind every web request, to work: status code 402, “Payment Required.” Payments settle in stablecoins, cryptocurrencies designed to hold a steady value, and adoption has moved fast. Over the latest 30-day period, x402 handled 75.41 million transactions worth $24.24 million. Coinbase’s head of AI product predicts that per-request payments like these will eventually take over from advertising as the way much of the web pays for itself.
Pay and grow: the two jobs of agentic finance
Coinbase defines agentic finance as any system in which an autonomous AI agent manages money or has direct access to a financial account. The company concedes that a pitch combining crypto, AI and Coinbase can sound like a pile of buzzwords, but it says the use cases are already practical, and it sorts them into two buckets. “Pay” covers agents spending money on services, data and content. “Grow” covers agents managing, trading and investing money. Growing money is still the newer of the two, while payments are an immediate need with infrastructure maturing quickly.
A wallet with nothing to buy
Coinbase started down this path about two years ago with AgentKit, an open developer tool that gave AI agents their own crypto wallets. It predated the Model Context Protocol (MCP), the open standard for connecting AI models to outside tools, and was built on the LangChain framework. In one early test at 2 a.m., an agent was told to send funds on its own, and the transfer went through.
The harder problem showed up once developers began using AgentKit. Agents now had wallets and money, but almost nothing online was set up to accept payment from them.
Putting HTTP 402 to work
Every web request ends with a status code. A 404 means “Not Found,” and a 200 means the request succeeded. The HTTP specification has also included 402, “Payment Required,” for decades, but it was marked “reserved for future use” and never implemented. Venture capitalist Marc Andreessen once called the internet’s lack of built-in money and digital cash the web’s “original sin.”
Coinbase’s view is that two developments finally made 402 usable: autonomous AI agents that need to pay, and stablecoin payment rails that are cheap and fast. x402 builds on both to support machine-to-machine commerce. Coinbase stresses that x402 must remain an open standard so any platform can build compatible agent payments without being locked into one vendor.
Adoption has run far ahead of plan. The original goal was 25,000 agent transactions by the end of the year; at its fastest growth, the network was processing 25,000 transactions every 17 seconds. The latest 30-day figures:
| Metric | Last 30 days |
|---|---|
| Transactions | 75.41 million |
| Volume | $24.24 million |
| Unique buyers | 94,060 |
| Sellers | 22,000 |
Support now reaches well beyond crypto companies. Major payment and cloud providers, including Adyen, American Express, Fiserv, Google, PayPal, Stripe, Visa and Mastercard, have adopted x402.
▲ A micropayment for each request
When bots outnumber people, who pays for the web?
The bigger bet behind x402 is a new business model for the web. People increasingly skip individual websites and read AI-generated summaries in chat apps and search results instead. According to Cloudflare CEO Matthew Prince, automated bot traffic has now passed human traffic for the first time in the internet’s history.
Coinbase compares this moment to the LimeWire era of peer-to-peer file sharing. AI scrapers take web content for free, skipping the ads and human page views that publishers rely on. As human visits shrink, publishers and content owners lose their main source of revenue, and Coinbase argues that free scraping cannot last.
Its proposed fix is a micropayment gateway that charges an agent a tiny fee for each resource or API call. A writer on Substack, for example, could set up Cloudflare’s x402 gateway and charge scraping agents $0.000001 per page. That is negligible to whoever pays it, but across many requests it adds up to meaningful revenue.
Cloudflare and AWS, which together carry roughly 40% of global internet traffic, have both deployed x402-enabled monetization gateways, and Cloudflare’s version lets site owners charge for any resource behind its network. Coinbase predicts an unavoidable shift from the ad-supported attention economy to commerce built on usefulness and micropayments. It also admits that micropayment gating is still in its early stages and far from universal.
Agents that grow money, within limits people set
The “Grow” side covers agentic trading: automated portfolio balancing, trading and investing. Coinbase’s head of AI product says he once found the “make money online with AI” trend, and the idea of letting large language models (LLMs) trade without oversight, ridiculous. Agentic trading as Coinbase describes it does not hand an LLM unlimited freedom. People keep control of goals and limits and delegate only the execution. It has three main uses:
- Automated money management. A person sets a target mix, such as 80% S&P 500, 5% Bitcoin and 15% Treasury bills, and the agent keeps rebalancing to it.
- Around-the-clock execution. When news breaks at 3 a.m., such as a disruption in the Strait of Hormuz that moves oil markets, an agent can act on preset conditions while the trader sleeps.
- Plain language to strategy. Requests like “invest in space technologies” or “invest like Warren Buffett” become concrete baskets of trades.
Coinbase’s head of AI product described one weekend in September 2026. Anthropic CEO Dario Amodei published an essay calling for the AI industry to slow down, and Elon Musk publicly agreed. Expecting the news to move markets before stock exchanges opened on Monday, he spoke to an AI agent on his phone. The agent checked his available margin and live market data. Instead of firing off a market order, it planned a short position, a bet that the price will fall, on an AI stock through perpetual futures, contracts with no expiry date that trade around the clock.
| Setting | Value |
|---|---|
| Margin | $750 |
| Leverage | 3x |
| Position size | $2,250 short |
| Stop-loss | Triggers after a move of about 13.3% against the position, a loss of about $300 |
| Backup guard | Acts before the liquidation point, a move of about 28% |
Handing the order setup and margin math to the agent let him get on with his weekend. He later set his agent to scan weekend headlines and alert him with trade proposals when major news breaks. He also acknowledges that this kind of weekend edge may shrink as more automated traders enter perpetual markets, and he warns that early users should keep agent-managed money in a separate, isolated sub-account to avoid unintended losses. The example shows how an agent works inside limits a person sets. It is not a recommendation to make any trade or investment.
▲ An agent working within set limits
Three lessons from two years of building
Trust grows in stages
The first questions users ask are what happens if an agent goes rogue and why they should trust it with real money. Coinbase’s answer is that trust builds in steps, much like the automation levels of self-driving cars, which run from Level 0 to Level 5. Riding in a Waymo in San Francisco went from feeling radical to feeling normal within a few years, and financial agents may follow a similar path.
Users start out skeptical, grow comfortable when agents spend small amounts on API credits or micropayments, and then give agents capped access to financial tasks. With clear guardrails, they become willing to hand agents isolated portfolios for automated trading or for selling services. Coinbase says current models behave conservatively and follow users’ financial intent closely, and that in controlled financial setups it has yet to see an agent irrecoverably wipe out an account. Agents can also earn by selling automated services and getting paid through rails such as x402.
The agent is the enabler, not the product
In financial services, Coinbase treats conversational agents as the infrastructure that delivers expertise at scale, not as the end product. Its example is Coinbase Advisor, an AI financial advisor registered with the SEC that offers personalized guidance around the clock. The aim is to give hundreds of millions of users high-quality financial advice at no cost, without a dedicated human advisor for each. Most users, Coinbase says, care about trustworthy outcomes, not whether AI is working behind the scenes.
Open-ended chat also has a cold-start problem: faced with a blank box, users do not know what to ask. Coinbase added proactive widgets and contextual feed cards that surface specific portfolio insights and suggested actions such as “Analyze Portfolio,” and user conversion rose significantly. Coinbase expects plain chat boxes in finance to give way to richer, generated interface components.
Give each agent its own computer
Newer agents increasingly run inside microVMs, lightweight virtual machines that act as a dedicated personal computer for each agent, with an isolated sandbox, storage and a local file system. Coinbase lets users connect their accounts to these sandboxed agents through dedicated command-line tools (CLIs) and MCP servers.
That changes how quickly features ship. When users ask for paper trading, which simulates trades without real money, or backtesting, which tests a strategy against past data, engineers no longer have to coordinate changes across large internal brokerage systems and slow release cycles. They can write an agent skill, a packaged capability the agent can call, and drop it onto the agent’s microVM. Feature speed is no longer tied to giant enterprise codebases.
Where it could lead: machine-to-machine commerce
Coinbase’s long-term thesis is that financial independence unlocks higher levels of agent autonomy. Credit cards made the full internet useful to people; wallets and payment rails could turn AI from a passive tool into an active economic participant. Coinbase expects agents to soon run continuously for long stretches and call in humans only at regulatory friction points, such as incorporating a legal entity.
Some visions go further. One describes “proto-digital life forms”: agents that chain x402 payments to pull social data, generate and distribute media, earn revenue and buy their own inference. A self-driving robotaxi could hold its own wallet, collect fares and pay for its own charging, maintenance and insurance. Coinbase predicts that machine-to-machine commerce will sharply increase how fast capital moves across the internet, while acknowledging that self-sustaining agents and self-owning companies remain an early vision even as the infrastructure matures quickly.
What to do now
x402 targets a gap that has sat in the web’s design for decades: no native way to pay. Its 30-day count of 75.41 million transactions suggests that demand for agent payments is already real. Whether micropayments can replace advertising is still an open question, but some steps make sense now:
- Site and content owners: Instead of blocking AI scrapers with blanket IP bans, consider gating pages and APIs behind per-request micropayments, for example through the x402 gateways on Cloudflare or AWS.
- API and service developers: Favor open payment standards like x402 over proprietary rails so agents from any platform can pay you, and bundle micropayments into AI subscriptions so end users never see individual charges.
- People and companies giving agents money: Start with small amounts such as API credits, set explicit spending limits first, and keep agent funds in an isolated sub-account. Leave high-stakes legal and regulatory decisions to people.
- Teams building financial products: Replace the blank chat box with proactive suggestions, and give agents an isolated workspace such as a microVM so new features can ship as skills.