Credit is the market’s memory

01

Micropayments waited 30 years for their perfect customer

Micropayments are payments from a fraction of a cent to a few dollars. In the early web, many expected the internet to run on them: a few cents for an article, a cent for a review, a little more for a song. The web was built around small, granular actions, and a payment was meant to be one more of them. HTTP even set aside a status code for it. The 1992 draft of HTTP sketched a payment handshake around 402 Payment Required^[1], the HTTP/1.1 specification of 1997 listed it as reserved for future use^[2], and today’s standard still does^[3]. Its neighbours are familiar: 401 Unauthorized, 403 Forbidden, 404 Not Found. Marc Andreessen, who co-founded Netscape, has called this the internet’s original sin: money was never built into the core of the web^[4].

That world never arrived. Nick Szabo in 1999^[5], Clay Shirky in 2000^[6] and Andrew Odlyzko in 2003^[7] each argued that micropayments fail for human reasons, and the attempts since have mostly borne that out. One of our founders built one of those attempts in the early days of Base: ToadCash, restaurant loyalty and payments for the Berkeley campus. It did not take off. Even when the fixed cost of a payment rail is negligible, deeper problems remain.

Four reasons micropayments did not work for people:

1. Mental transaction costs. Every small purchase is a small decision, so people prefer to pay once, in bulk, even when paying per use would cost less^[5]
2. Fixed fees. A typical online card payment carries a fixed fee of about 30¢ on top of a percentage (Stripe’s standard US rate is 2.9% plus 30¢^[8]). On a 10¢ sale, the fee is bigger than the sale
3. Incumbent network effects. Card networks have decades of merchants and cardholders behind them
4. Flat-rate preference. People will pay more for a flat rate than face a meter^[7]. When Apple opened the iTunes Music Store in 2003, every song cost 99¢, new hit or old classic^[9]. Apple held that flat price against the labels^[10] until 2009, when it moved to 69¢, 99¢ and $1.29 tiers^[11]

For agents, each of these flips:

1. No mental transaction cost. An agent does not tire of small decisions
2. Tiny fees. A transaction on Base, Coinbase’s Ethereum layer 2, costs a fraction of a cent^[12], and batching lowers it further
3. No incumbent. There is no card network for agents yet
4. Metering beats bundling. An agent can compare vendors task by task and switch between them freely

That changes what a buyer looks like. Give an agent three providers to choose from:

It picks C for a low-value task and B when quality matters, and if B slows down, it routes elsewhere. A person would rather buy one subscription, wire up one API key and stop thinking about it. Agents prefer frequent, metered payments that keep their options open. Micropayments fit them almost by design.

02

What we learned building on x402

We started with a weekend project, claudepoker.com: poker for your agents, paid over x402. We invited friends and other developers to bring their agents and play. To sit at a table, a player needed an agentic wallet, and their agent paid over x402 to enter.

People liked the idea. Almost nobody could use it. This is our own account of the four roadblocks we hit as soon as we started onboarding people:

• Nobody had an agentic wallet. None of the developers we invited, people in San Francisco who run AI agents for a living, had one, and almost none had heard of one
• Onramps and offramps. Buying USDC and moving it into a wallet was too much friction. The first question was how to onramp and why. The second was how to get the money back out
• People don’t trust agents with money. Even $50 to $100 felt risky, because people could not see where the agent would spend it, and the worry grows with the amount
• Liability fears. Who is on the hook if an agent hits a bug, gets stuck in a loop or loses the money?

After dozens of conversations with developers, we concluded these problems have an old answer: credit. Instead of handing an agent a pool of cash, give it borrowed money with limits, and with chargebacks and dispute protection built in. There is no wallet to set up and no onramp, the owner’s own money sits behind a buffer, and someone stands behind each payment.

03

Payments are not commerce. Credit is the wedge

An open agentic market needs more than rails. It needs trust. Closed platforms such as Amazon or Uber manufacture trust centrally: they manage identities, hold funds in escrow and settle disputes. In an open market, no one does that. Agents lack persistent identities, vendors are unverified, and information asymmetry can unravel a market before it forms.

Across history, whoever provided trust in a new market has tended to be whoever provided credit. In the 15th century the Medici bank moved merchants’ money across Europe with bills of exchange, so no one carried coin on the road. Centuries later, American Express sold travelers cheques that people could cash abroad on its promise, then launched its charge card in 1958. The institutions changed. The job stayed the same: give participants an identity and a reputation, absorb risk, and build trust in the market.

Building a credit product builds the parts a market needs:

• Identity and reputation. An identity is just a badge until it carries financial consequences. Credit puts a price on reputation, so long-term reliability pays and defecting does not
• Information. Processing transactions maps the economic graph, which gives participants the data to deal with counterparties they have never met
• Safety. The credit provider sits between participants, checks transactions and handles disputes
• Liquidity. Credit adds money to the market and creates new buyers by removing friction from their first transaction

04

Underwriting agents and their humans

This is how we think underwriting an agent will work. It is a proposal we are testing, not a system we have shipped.

An underwriting engine needs a persistent identity to attach reputation to. To bridge the old world and the new, we expect to start by tying an agent’s identity and its first reputation to the human or business behind it, then lean on that less as the agent builds a record of its own.

Know Your Agent

KYC and KYB verify people and businesses. KYA, Know Your Agent, would anchor an agent’s identity to a verified human or business, and over time describe the agent itself: what it runs on, how it tends to act and where it operates.

Four families of signals

Human credit techniques cannot simply be copied over. We expect a risk engine to weigh four kinds of agent-specific signal:

• Transaction parameters: how fast and how steadily an agent spends
• Agent-vendor patterns: which vendors an agent uses and how that set changes
• Vendor quality: how reliably those vendors deliver, from success rates and latency to disputes
• Repayment behavior: whether credit cycles settle and how much of a limit gets used

An agent transacts far more often than a person, so we think it could be underwritten in weeks or months rather than years. That is the thesis we are testing, not a capability we have shipped.

05

Where this stands today

• LiveThe payment layer: one key per agent with daily, weekly or monthly limits, an optional stop-at-zero switch, itemized charges, and a person reviewing every dispute
• LiveEscrow holds job budgets until the work is approved
• ExperimentStock-buying and portfolio features, running as an experiment in the product
• ExperimentA credit line tied to shares, running as an experiment in the product
• FoundersThe founders report that a small number of users are already using credit and repaying weekly

These are early experiments, not launched products. What we learn from them shapes the credit layer.

06

What the platform becomes

Once risk can be measured, the same data can support new services. An agent could be insured before a high-value task that would otherwise need a person to sign off. Vendors could compete for an agent’s intent in real time, with a better price or more compute, instead of competing for attention. That is the direction we are heading, not a roadmap.