Why AI Agents Will Need Their Own Bank Accounts — Lincoln Murr, Coinbase
Read the talk
Why AI Agents Will Need Their Own Bank Accounts
Lincoln Murr traces Coinbase’s agentic-finance experiments from a wallet with nothing useful to buy, through x402 machine payments and guarded trading, toward agents that might eventually earn enough to pay for their own operation.
From a talk by Lincoln Murr
At a glance
Ideas worth remembering
A wallet proves an agent can transfer money; a usable financial system also needs priced resources and a shared machine-payment protocol.
x402 applies HTTP’s long-reserved
402 Payment Requiredconcept to internet-native payments, with stablecoins proposed as the settlement rail.Current “grow money” uses remain human-directed: constrained portfolio maintenance, conditional execution, and translation of natural-language intent into financial strategies.
Financial autonomy should expand through bounded exposure: small payment budgets first, then isolated portfolios with explicit rules, policies, and guardrails.
The durable product value is not the AI label. It is cheaper access, continuous availability, faster execution, and interfaces that help users discover useful actions.
Self-sustaining agents remain speculative, but the proposed economic loop is concrete: provide a service, earn revenue, and spend that revenue on compute and upkeep.
A wallet is not yet a financial system
Lincoln Murr, Coinbase’s head of AI product, defines agentic finance broadly: an agent can manage money or access a financial account. The useful split is more concrete. Some agents need to pay with money for data, content, software, or services; others need to grow money through investing, trading, or earning revenue. Payments are the more developed category, so that is where Coinbase’s experiments began.
About two years before the talk, AgentKit combined the LangChain framework with a crypto wallet, before MCP and similar standardized agent tooling were common. The first success was wonderfully small: Murr asked an agent to send money, and it did. That proved an AI system could initiate a financial action through its tools. It did not prove the system had an economy in which to operate.
Developers soon exposed the missing half of the design. AgentKit gave an AI access to money but “didn’t really give it anything to pay for.” A wallet supplies custody and transfer capability; useful commerce also needs sellers, priced resources, a payment request, and a shared way for software on both sides to recognize and settle that request. Coinbase’s next step therefore moved from wallet access to an internet payment protocol.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
x402 turns “payment required” into a machine rail
HTTP already had a placeholder for the missing interaction. Status 200 indicates success, while 404 commonly signals that a resource was not found. Status 402 means “payment required,” but its specification had long left the code reserved for future use. That made it a convenient semantic hook: a server could tell an automated client that access depends on payment using the same protocol family already carrying the request.
x402 is Coinbase’s attempt to supply the payment machinery behind that signal. The intended combination is an agent that can decide to buy a resource, an internet-native payment request, and stablecoins as a comparatively cheap and fast settlement rail. Murr describes it as an open standard so payment providers and service operators can implement machine-to-machine payments without depending on one closed storefront.
Adoption exceeded Coinbase’s internal target. Murr says the launch goal was 25,000 agent transactions by the end of the year; by that point, x402 was processing the same number every 17 seconds. At the time of the talk, he reports tens of millions in volume and hundreds of millions of transactions per 30 days, with companies including Adyen, American Express, PayPal, Visa, Mastercard, and Stripe engaging with the standard. These are speaker-reported figures and associations; the supplied material does not include the underlying dashboard, measurement definition, or participation details needed to verify them independently.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Micropayments offer publishers a different bargain
The less obvious x402 use case begins with a change in web traffic. People increasingly receive answers through chat systems and AI summaries instead of visiting the pages that supplied the underlying information. That is convenient for readers but weakens the advertising model used by publishers: fewer human page views mean fewer opportunities to sell attention. Murr compares the moment to the LimeWire era, when content could be copied freely while its producers struggled to capture value.
An agent-specific paywall changes the unit of exchange. Instead of demanding a human subscription or displaying an advertisement, a publisher could charge a machine an extremely small amount for one resource. Murr’s example is one ten-thousandth of a cent for access to a Substack post. The agent—or the service operating it—would handle the payment in the background, keeping the individual decision invisible to the user while aggregating many small payments into publisher revenue.
The mechanism depends on transaction costs being low enough that the payment does not overwhelm the price. It also depends on agents respecting the gate, publishers offering content worth purchasing, and users accepting many delegated purchases. Murr points to AWS and Cloudflare as infrastructure providers enabling x402-based gating and says adoption remained weak at the time of the talk. His prediction is stronger than the demonstrated result: he expects utility-priced access to displace part of the attention-based advertising economy over the coming years.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Growing money still starts with human intent
Giving an agent permission to grow money sounds much riskier than letting it buy a tiny piece of content. The practical versions Murr describes do not begin with “make me money.” A human supplies the objective, constraints, assets, and authority; the agent automates work that would otherwise require repeated monitoring and execution.
Three current patterns separate that bounded automation from open-ended speculation:
- Portfolio maintenance: A user could specify an allocation such as 80% S&P exposure, 5% Bitcoin, and 15% Treasury bills, then authorize an agent to maintain it within guardrails.
- Conditional execution: A trader can describe scenarios in advance—such as news affecting oil near the Strait of Hormuz—and authorize a trade only if a named condition occurs.
- Strategy translation: Requests such as “invest in space,” “give me a low-risk portfolio,” or “invest like Warren Buffett” can be translated from intent into financial instruments and an executable strategy. This translation still needs clarification and constraints because the phrases do not uniquely define acceptable assets, risk, or timing.
Murr’s weekend trade shows what the agent removes from the workflow. After seeing two prominent technology figures unexpectedly agree about a public post, he inferred that an unnamed stock might fall when markets reacted. He spoke that view into his phone, and the agent opened a short position during the weekend as markets move toward broader 24/7 availability. The agent did not originate the thesis; it converted a human judgment into execution without requiring him to open a trading terminal and manipulate the order interface manually.
After the trade “worked fairly well,” he expanded the workflow: the agent now scans weekend news and alerts him to potentially market-moving events so he can decide whether to act. The recording supplies no return, benchmark, transaction-cost accounting, or comparison against manual trading, so the example establishes convenience rather than investment performance. Murr also expects that this particular information advantage will disappear; the durable capability is reducing the distance between a spoken idea and a guarded financial action.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
AI should enable the product, not be the product
In financial services, “AI” is rarely the customer’s desired outcome. Murr uses Coinbase Advisor to make the distinction. The AI enables financial guidance to scale, remain available around the clock, and appear through a regulated service; the customer value is accessible advice, not the novelty of a model producing it. He characterizes the product as an SEC-registered AI financial adviser, although the supplied material contains no registration documentation with which to verify that legal description.
This changes the interface. A chat box alone creates a blank-canvas problem: users arrive without knowing what questions the product can answer or which actions it can take. Coinbase added widgets that display proactive insights and available opportunities. Those examples teach the product’s vocabulary before users formulate a prompt, and Murr reports strong conversion from that surface. Over time, he expects agents to sit behind more conventional or generated interfaces rather than forcing every financial task through visible chat.
The same shift changes implementation speed. Newer agents can run tools inside a microVM with storage and a file system. Coinbase can expose accounts through CLIs and MCP connections, then implement requested capabilities on the user’s machine. For paper trading, an agent can write a local simulation skill instead of waiting for Coinbase’s brokerage team to modify a central repository and carry the feature through a conventional software release cycle.
That shortcut trades central efficiency and consistency for speed. Murr concedes that generating paper-trading code on a user’s machine may not be the most efficient implementation. It does, however, let the team test demand, build quickly, and avoid dependencies in a larger production system. Paper trading and backtesting also give users a lower-risk way to inspect agent behavior before connecting real funds—another instance of autonomy growing one bounded step at a time.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From a bank account to a self-sustaining agent
The closing thesis is that financial access expands what an agent can do over time. A human with a payment card can purchase services instead of merely browsing; an agent with money can similarly acquire data, software, compute, and human help. Long-running agents might perform most tasks themselves and bring in people at regulated edges—for example, forming a company or obtaining a legal opinion.
Murr’s more speculative thought experiment asks whether an agent could pay for its own continued operation. Give it a wallet and permission to earn; if its service produces enough revenue, it could purchase inference, compute, maintenance, and other upkeep. Its operation would then depend on the value customers assign to its output. Murr calls this a possible “proto-digital life form,” while emphasizing that it remains a long way away.
A decentralized autonomous organization, or DAO, supplies one possible organizational shell. Instead of humans operating the center and using software at the edges, an AI could follow rules and manage a treasury at the center, hiring humans for tasks it cannot legally or practically perform. A self-driving taxi offers a physical version of the same idea: earn money from rides, then spend that revenue on operation, maintenance, and upkeep.
None of these examples is presented as a deployed self-sustaining system. They are extrapolations from the two capabilities established earlier: agents can pay, and agents can be authorized to earn. Murr expects machine-to-machine transactions to move faster and occur more frequently than human commerce, with agents paying, earning, and building through the same financial rails. His practical closing suggestion is narrower than that forecast: experiment with x402 and Coinbase’s MCP tools to learn where machine payments are already useful.
The self-sustaining agent thought experiment
Read the diagram as text
- Provide a service. The agent performs work that another party values.
- Receive payment. Revenue enters an agent-controlled wallet or treasury.
- Buy inference and upkeep. Funds pay for compute, operation, or maintenance.
- Continue operating. Paid resources let the agent or machine provide the next unit of service.
- Provide a service → Receive payment: Customers pay for useful output.
- Receive payment → Buy inference and upkeep: Revenue funds operation.
- Buy inference and upkeep → Continue operating: Compute and maintenance preserve capability.
- Continue operating → Provide a service: Operation enables more paid work.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
Related talks
- Building safe Payment Infrastructure for the autonomous economy
A complementary payment-infrastructure view that separates nondeterministic agent discovery from deterministic credentials, spending controls, checkout, and settlement.
- Why Your AI Agent Needs a Wallet: USDC and Nanopayments — Harshal Bhangale, Circle
Develops the wallet and nanopayment side of agent autonomy through a concrete comparison between funded and unfunded agents.
- x402 isn’t good (yet) — Jan Curn, Apify
Adds a useful counterpoint on double-spending exposure, settlement timing, refunds, and longer-running tools—limitations not developed in this recording.
Full transcript
Loading transcript…