HTTP 402 Slept for Thirty Years. Coinbase Turned It Into a Payment Protocol for AI Agents

Nathan Labenz opens with a status code almost nobody has implemented. HTTP 402 Payment Required has existed since the early web, named as a placeholder for a future in which a site could demand a few cents before returning a page. Visa and Mastercard were never going to clear a one-cent charge; interchange would have cost more than the sale. The internet filled the hole with ads and subscriptions instead.

His guest is Nemil Dalal, developer platform lead at Coinbase and a co-creator of x402, an open standard that reuses that dormant code. A seller adds a library, returns 402 with a price, chain, and address, and waits. A buyer — a person or an AI agent with a crypto wallet — pays and retries with proof attached. Labenz's premise is that agents are now reliable enough that handing them a modest budget for APIs, MCP servers, or even other agents is no longer a thought experiment, while card networks remain built for humans and priced for larger tickets.

Dalal used to run USDC at Coinbase when it was a venture bet of tens of millions in market cap, and he helped grow it to a billion. Labenz admits he is crypto-curious rather than crypto-fluent: he has sent USDC through Coinbase to international contractors and found the surface magical, without knowing what sat underneath. Before they design the agent economy, he wants a stablecoin 101, then the protocol, then the harder questions about wallets, reputation, and what happens when models can spend.

In about a decade, Dalal says, the world went from zero stablecoin usage to roughly $250 billion, almost all dollar-denominated, with Coinbase and Circle among the largest via USDC. Blockchains are globally reachable, often cheaper than bank rails, and settle in seconds. Fiat's gift is that you can buy groceries in it. A stablecoin is the attempt to keep both properties at once, so sending dollars feels as simple as sending a text.

The peg is not a mystery algorithm. The issuer takes dollars, parks them in a bank account or short-dated government paper, and mints tokens one-to-one: $10 in, 10 USDC out; redeem and they burn the tokens and return fiat. That is the opposite of Bitcoin. "You absolutely have to trust someone," Dalal says — Circle and Coinbase as issuers, plus whoever holds the mint-and-burn keys, because those keys must not be able to print USDC without reserves. The firms publish transparency reports; Congress is exploring legislation that would put a legal wrapper on that trust.

A concrete path: ACH a few cents and two to three days from a U.S. bank into Coinbase, mint 10 USDC, and send to anyone with a wallet — they need not be a Coinbase customer. A recipient in India can hold USDC, put it in DeFi, or sell to a market maker and cash out to rupees, paying perhaps half a percent to a percent. Dalal, who is Indian, notes you still cannot spend USDC at every local shop. The destination he wants is a closed loop in which an onchain dollar or rupee is simply better money: cheap to send, globally usable, and plugged into the rest of decentralized finance.

The business is the float. At about 3 to 4.5 percent on $250 billion, issuers are already making more than $10 billion a year, money that can pay market makers, cash-in and cash-out desks, branches, and compliance. Labenz asks whether this is an Uber-style subsidy era, burning venture cash to get people "onchain." Dalal says no: these are already great businesses, closer to a bank spread with a cheaper cost structure. Coinbase shares some of that yield with users who keep USDC on the platform; park the same tokens in a self-custody wallet for a year and Coinbase and Circle keep it.

Dalal's other punchline is price. A wire still costs real money in a world where an SMS is free, even though much of what moves is information rather than vaults of cash. In 2017 a typical onchain transfer was $2 to $5; when gas spiked it reached $100 to $200. On Base and similar networks today it is a matter of cents, and sometimes less, without a subsidy. He spent early years around banks still running 1950s-to-1970s stacks. Convincing those institutions to change is hard; a new rail built like the internet just is cheaper.

Labenz asks about the other side of that bargain. Cards buy chargebacks and fraud desks; crypto looks cheaper because it is not reversible by default. Dalal agrees on the core. Exchanges still run KYC. USDC issuers work with governments and can freeze funds when law enforcement has a case, including U.S. sanctions. You cannot press a button and unwind a payment, which is also why the recipient can trust a two-second credit instead of waiting days. Reversibility can be layered on later — products exist and are not yet popular. Fraud costs explain some bank fees, he says, not a $20 or $30 wire when a chain can move a trillion dollars for less than a cent.

The collapse in cost has a 101. Dalal analogizes to Moore's law for bandwidth: Bitcoin launched in 2009, and the industry has had sixteen years. The largest shift is more block space. In 2017 an Ethereum transfer sat on layer 1; now layer 2s, and even layer 3s, execute the activity and post a summary back, closer to how banks net rather than settle every swipe. Base recently hit almost 1,000 transactions per second, with sub-500-millisecond settlement on the roadmap.

The other lever is a trade-off Bitcoin refused. If every human must run a node on a phone, capacity stays low; if thousands of operators can run serious nodes on AWS, throughput jumps. Different networks chose different points on that curve. Labenz hears a mix of raw efficiency and a quieter walk-back of "everyone runs a node." Dalal says the whole industry is mixing those approaches, and he expects people to transact onchain the way they moved onto the internet — not only payments, but social networks and the rest of the stack.

Coinbase shipped AgentKit so an agent could hold a wallet. Thousands of developers picked it up, including teams such as Virtuals that launched tippable AI personalities. The next request was obvious: let that wallet buy a good, hit an API, or send a text. Credit cards never fit. They need human permissioning, and an agent does not subscribe to one service — it goes to A for one query and B for the next. The 1990s web imagined 402 as "pay ten cents for this page" and could not get the card networks to play along. x402 revives that idea as an open standard anyone can build on.

Two sketches: a storefront that takes 20 USDC and ships to an address the agent provides; an API that sells market data, inference, or Twilio-like messaging by the call. Dalal thinks humans will use it for micropayments too. The urgent case is agents and MCP servers, because those clients are not watching ads. MCP made tools discoverable on a web built for people; it still lacked a payments layer. x402 is that layer. Documentation, a white paper, and an ecosystem list live at x402.org. They had launched only a few weeks before the taping, starting with crypto APIs and talking to inference providers.

The loop he walks through is inference, a use case close to home: an agent that needs more compute hits an API and gets 402, with a price, a network such as Base or Solana, and a destination. It signs a transaction — five USDC on Base, in the example — and sends that payload with the retry. It does not broadcast. The provider checks that the transaction is well-formed and that the money is there, does the work, then decides whether to submit. Overpay can become a prepaid balance; underpay gets a "send more" response instead of the result.

That is the whole protocol: request, signed payload, verify, work, settle, respond. Middleware for the seller is one line of Node.js in front of an existing API, but a standard on paper is not enough. Coinbase Developer Platform, which Dalal leads, will spin up a merchant wallet and auto-convert incoming USDC into dollars in a bank account. On the agent side, a 402 becomes another tool: pay, then loop back. Labenz asks whether KYC at labs such as Anthropic and OpenAI collides with wallet rails. Dalal treats that as the model's independent product choice. x402 is the lowest-level rail; a reputation or credit layer for an agent that is short on cash can sit on top later, the way credit networks sit on cards. That layer does not exist yet.

Dalal's larger claim is that the web's monetization model is going to change. Agents do not sit through display ads, so if more traffic is agent traffic, the side of the internet that sold attention has a problem. He floats a future in which a publisher such as the New York Times charges an AI crawler at the door: no payment, no page. Labenz finds the human version intuitively appealing too. Sidebar ads do not mint cash, subscription stacks are exhausting, and a reader might pay one to ten cents for a CNN piece without another $30 year.

Why did this not already happen, given that blockchains have existed for years? Traditional cards still cannot charge a cent. Early crypto had no stable unit of account, so you could not price an article in dollars. Fees on Bitcoin ran $5, $10, $20, even $30; even $2 killed a one-cent purchase. Only in the last six to twelve months, Dalal says, have Layer 2s, Solana, and others brought costs down far enough. Wallet addresses were also hostile. A few crypto-native apps such as Zora and Farcaster already charge a few cents. The mass web could not.

Labenz hears the AI story in that timeline. People ask why capable models are not everywhere; his answer is that they were not useful until about two years ago, context windows were 8,000 tokens, and they were not multimodal. The thresholds for real usability were only recently crossed. Dalal likes watching the two journeys rhyme. Crypto is older, but only in the last few years did it get simple enough for a mass market. Six months of crypto or AI, he says, can make the previous six months look unrecognizable.

Labenz cannot draw a clean box around "the agent." A past guest, Augment, shipped an enterprise coding assistant that called a smart MCP from another open-source developer to plan code changes. If that MCP charged, who is the agent, and who is at fault for a bad plan? Dalal sees generalists calling specialists. Coinbase Developer Platform already has AIs paying other AIs, which still surprises people outside the field. Ads will not fund that path. If it is one company's stack, billing can be internal; across companies and context windows, you need a financial layer.

He also wants the chain as an identity layer. An NFT can be an empty passport; each confirmed good act — correct medical advice, say — is a stamp from someone who checked the work. Readers weigh who stamped it, like an open-source credit bureau. One bad review does not erase 999 good ones. None of this is science fiction, he says. The unsolved part is which use case to start with, and he would like entrepreneurs in the audience to try.

Labenz's first instinct is a $10 or $100 agent wallet he can afford to lose to prompt injection. Dalal agrees segregation is the start, then says people will put millions, if not billions, into agents over time. A trading or wealth-management agent needs real size, and hundreds of people at a Coinbase hackathon were already building versions of that. Tyler Cowen's idea that agents should be capitalized maps onto crypto slashing: stake in escrow, earn if you behave, lose the stake if you do not. The jury can be a community, elected leaders, or a trusted centralized attester. The smart contract still executes the same way worldwide, which matters in a world that does not share one legal system.

Spending rules belong in the wallet, not the system prompt. Dalal describes a 2-of-2 multisig like a safe-deposit box: the agent can spend up to $10 alone; above that the human's key must co-sign, or the transaction fails. You still tell the agent to ask, because a good agent will prompt you — you just do not trust that. On systemic risk, Labenz worries about a flash-crash on a chain nobody can halt. Dalal says the agent-wallet economy is still tiny, maybe a handful of wallets before AgentKit. The practical brakes are reputation, a human in the loop, and exchange-level circuit breakers on venues such as Uniswap or Aerodrome if a single block runs too hot. When amounts are small, people still look. When the alerts feel like popup notifications, they will abdicate, and protocol constraints become the last defense.

Labenz has been looking for someone to put the smart in smart contracts. Code is brittle. It cannot sit as a human judge over a fight about a hundred-year-old house repair. He has not seen an onchain product that calls Claude or Gemini on submitted evidence and releases funds. Dalal has not either, and he wants a listener to build it. Architecturally it is a multisig in which one key is the model: two-of-three, two humans disagree, the model casts the tie-breaking vote.

Earlier in his career he spent a few months with the Gates Foundation on weather-index crop insurance. If rainfall is below a threshold, farmers are paid automatically; the hard part was deciding whether the season qualified. An AI could review precipitation over six months and unlock the escrow. In crypto this is the oracle problem. Contracts see only onchain state; weather lives off-chain. Using a trusted model as that oracle is, he says, relatively simple today. Anyone who wants to try can reach him on Twitter.

Labenz immediately flags gullibility. Models are easy to trick. A photo of a roof might be a month old; a field sensor might not be where it claims. Provenance — device-signed photos, attested sensors — still looks thin. Dalal splits the work in two: is the evidence authentic, and is the model hallucinating or being goaded? Community agents with Twitter accounts have already been talked into sending money they should not. Start with satellite rainfall, not roofs. Keep the model off open chat. Do not begin with the hardest dispute on day one.

They are still scratching the surface of crypto as a financial system for agents. x402 is the start; reputation has barely been tested. Builders should try the white paper and middleware at x402.org. Labenz's larger reminder, from the top of the show, is how contingent the AI stack already is. GPUs were a video-game accident that enabled the wave. Cryptocurrency, designed so nation-states could not seize the money, may be the accident that lets models join the economy — which is a reason to be humble about forecasts, and a reason to invest now in the monitoring and control layer around agent wallets.