You're building agents that compose across providers. The billing question is coming, and the answer might be simpler than everyone thinks.

Five people at dinner. The bill arrives. Three options: one person pays and everyone settles up later, split it evenly, or itemize every dish down to the garnish. The most accurate method is itemizing. Almost nobody does it. The overhead isn't worth the precision.

I've been thinking about this mundane little problem since writing about how every action is becoming an agent.¹ When your email agent calls my calendar agent, which negotiates with their scheduling agent (a currently feasible scenario), the composition works beautifully. MCP provides the shared language. Google's A2A protocol handles discovery.

But when the task is done and five agents from three providers each contributed a piece of the result, someone has to split the bill. And the tech industry, predictably, is reaching for the most precise method possible.

The Precision Trap

Four separate agent payment standards appeared in under a year.²

Coinbase revived the HTTP 402 "Payment Required" status code (reserved in the original HTTP spec for future payments but never used) to settle agent transactions in real time using digital currencies. Google shipped an Agent Payments Protocol with digitally signed payment instructions. OpenAI and Stripe co-developed a commerce protocol for agents buying things on behalf of users. Google and Shopify launched another standard with Walmart, Target, Best Buy, and Macy's on board.

The ambition is impressive. Real-time settlement. Per-transaction micropayments. Fractions of a cent tracked and cleared instantly.

It's also, I think, solving the wrong problem.

How Economies Actually Settle

If you're building agents that might eventually charge for their services, consider how other multi-party economies handle the same question.

When I call someone on a different phone network, my carrier doesn't pay their carrier in real time. They track the traffic, batch it monthly, net the bilateral balance, and one carrier sends the other a wire transfer for the difference. Billions of calls. One monthly payment.

When a song plays on Spotify, the songwriter, performer, producer, label, and publisher all have a claim on fractions of a cent. Spotify doesn't make six payments per play. It aggregates plays over a quarter, calculates shares according to agreements, and distributes in batches.

Financial markets invented the clearinghouse for exactly this reason. London bankers' clerks in the 1770s started meeting at a tavern on Lombard Street to exchange notes and settle mutual debts. The insight that stuck: instead of every party settling with every other party, a central counterparty nets all obligations and settles the differences. Bilateral chaos becomes a single balance.³

The pattern is the same everywhere I look. In every multi-party economy, the settlement system that wins optimizes for low overhead, not high precision. Track usage. Batch it. Net it. Settle the difference. Move on.

Why Micropayments Keep Failing (Even for Machines)

This pattern explains something that's been nagging me about the agent payment conversation.

The internet tried per-transaction micropayments before. DigiCash went bankrupt in 1998, reportedly after turning down a deal to embed electronic cash in every copy of Windows 95.⁴ Beenz raised $86 million and folded. Flooz got Whoopi Goldberg to do TV ads, then collapsed when criminals exploited the system with stolen credit cards.⁵ The W3C formed a micropayment working group that defined HTML tags like <price>. It closed in 2001 without adoption.⁶

The standard explanation is that humans hate payment friction. True. But it misses the deeper problem. Per-transaction micropayments create accounting overhead that's disproportionate to the value of the transaction. A payment system that tracks, validates, clears, and settles every $0.003 interaction individually is burning energy on precision that nobody needs, regardless of whether the payer is a human or a machine.

The crypto counter is real: on modern blockchain networks, a payment settles in under a second for a fraction of a cent. The transaction cost problem is genuinely solved. But the overhead I'm talking about isn't the payment fee. It's the bookkeeping. Tracking, validating, and reconciling millions of individual transactions across providers. Telecoms could technically settle every call individually. They batch it anyway, because the operational complexity of per-transaction accounting isn't worth the precision.

The internet didn't solve its payment problem with micropayments. It solved it with aggregation. Your cloud hosting bill doesn't list every API call individually. It batches usage and charges once a month. Stripe doesn't settle each purchase in real time. It batches daily and deposits the net amount. The infrastructure that actually won is boring, periodic, and aggregate.

The Boring Prediction

So here's where I land.

The agent economy's payment infrastructure won't look like a cryptocurrency exchange. It won't require anyone to understand digital wallets or novel financial instruments. It'll look more like a phone bill.

Agent providers will track inter-provider usage in a shared ledger. Periodically, a clearinghouse function will net the bilateral balances. Provider A's agents used $4,200 of Provider B's services this week. Provider B's agents used $3,800 of Provider A's. Net settlement: a $400 bank transfer. No exotic payment rails. Just usage tracking, netting, and a wire.

The interesting engineering isn't in the payment mechanism. It's in three things that sound mundane but are genuinely hard:

  1. Usage metering. When a chain of agents from multiple providers produces a result, who contributed what? This is an attribution problem, and it's similar to what the advertising industry has struggled with for decades. Agents will need to emit standardized usage records, the way telecom networks emit call detail records.
  2. Service-level agreements. What happens when an agent in a five-agent chain fails or returns garbage? Telecom has interconnection agreements that specify quality standards and dispute resolution. Agent providers will need something equivalent. Boring contracts, not exotic technology.
  3. A universal service catalog. Before you can bill for a service, you have to describe it in a way both sides agree on. MCP already describes what an agent can do. The missing piece is attaching a price to each capability. Something like a tariff table: "translation, per 1,000 words, $0.05" or "data enrichment, per record, $0.002."

None of this requires inventing new money. It requires inventing new accounting.

What This Means If You Build Agents

If you're an indie hacker building an agent that does one thing well, the payment question might feel distant. But it's closer than it looks.

The protocol layer (MCP, A2A) is making it possible for your agent to be discovered and called by other agents across providers. The moment that works reliably, the payment question goes from theoretical to urgent. And the builders who've already thought about it, who've attached a price to their agent's capabilities and can emit clean usage records, will be ready to plug into whatever settlement infrastructure emerges.

You don't need to pick a payment protocol today. You need to be able to answer one question: what does your agent charge per unit of work? The plumbing for collecting that charge will come. What won't come automatically is clarity about what the unit is and what it costs.

Back to the dinner table. Five agents, one result, three providers. The settlement method that'll win won't be the one that tracks every fraction of a cent in real time. It'll be the one with the least overhead that everyone can trust.

The most consequential infrastructure is usually the most boring.

If this line of thinking resonates, you might also enjoy:

  1. "Every Action Is an Agent," my earlier post on how application boundaries dissolve when every software action becomes an autonomous agent connected through MCP.
  2. Between May 2025 and January 2026: Coinbase launched x402 (HTTP-native agent payments using USDC, a dollar-pegged digital currency, on the Base blockchain), Google shipped AP2 (Agent Payments Protocol extending A2A and MCP, with Mastercard, American Express, and PayPal among 60+ partners), OpenAI and Stripe co-developed the Agentic Commerce Protocol, and Google and Shopify launched the Universal Commerce Protocol (with Walmart, Target, Best Buy, and Macy's).
  3. Clearinghouse mechanisms emerged when London bankers' clerks began meeting at the Five Bells tavern on Lombard Street around 1770 to exchange notes and settle mutual debts. The first dedicated Bankers' Clearing House was built on Lombard Street in 1833, funded by 39 founding bankers including Barclays, Glyn, Martin, and Williams. Central counterparty clearing, where one entity nets all obligations, reduced settlement risk across every major financial market that adopted it.
  4. DigiCash was founded by David Chaum in Amsterdam in 1990, based on his 1982 blind signature cryptography. The Bill Gates offer is widely reported but unverified from primary sources. Only one US bank (Mark Twain Bank in St. Louis) adopted eCash, signing up 5,000 customers in three years. DigiCash filed for bankruptcy in 1998.
  5. Flooz raised $35 million and enlisted Whoopi Goldberg as spokesperson. After the FBI discovered a Russian-Filipino organized crime ring using stolen credit cards to buy $300,000 in Flooz currency, the company's credit card processor froze its accounts, triggering a cash flow crisis. Beenz raised $86 million from Softbank, Apax Partners, and Vivendi. Both shut down in August 2001, also casualties of the dot-com bust and the chicken-and-egg problem.
  6. The W3C Micropayment Markup Working Group (1999-2001), chaired by IBM and Compaq representatives, defined markup for embedding payment information in web pages, including standardized price fields. It closed in June 2001 without adoption.