The fire-hydrant security system

Six trillion dollars a year was protected by a photograph of a fire hydrant. That worked. Right now a machine is trying to buy something. Four different companies disagree on whether it should be allowed to. $8 billion has been deployed on this question. The machine doesn't know this. It's still trying.

In 2000, a Carnegie Mellon professor named Luis von Ahn invented the security system that would protect internet commerce for the next fifteen years.

The system required you to look at a grid of blurry photographs and click the squares containing a fire hydrant. If you clicked correctly, you were considered human. If you were considered human, you received the credit card form.

He won a 𝗠𝗮𝗰𝗔𝗿𝘁𝗵𝘂𝗿 𝗚𝗲𝗻𝗶𝘂𝘀 𝗚𝗿𝗮𝗻𝘁 for this. $800,000.
Presumably in a check he was not required to identify traffic lights to cash.

That fire hydrant stood between you and six trillion dollars in annual commerce.

Then AI learned to solve CAPTCHAs in 0.3 seconds, so the hydrant is retired, and 𝟭𝟲𝟮 𝗽𝗿𝗼𝗷𝗲𝗰𝘁𝘀 𝗮𝗿𝗲 𝗻𝗼𝘄 𝗱𝗲𝗽𝗹𝗼𝘆𝗶𝗻𝗴 $𝟴 𝗯𝗶𝗹𝗹𝗶𝗼𝗻 to build the exact opposite system: infrastructure that decides not whether the buyer is human, but whether the machine is authorized.

The gate is still there. It just faces the other direction.

𝗠𝗼𝘀𝘁 𝗰𝗼𝘃𝗲𝗿𝗮𝗴𝗲 𝗶𝘀 𝗮𝗯𝗼𝘂𝘁 𝘁𝗵𝗲 𝗽𝗿𝗼𝘁𝗼𝗰𝗼𝗹 𝘄𝗮𝗿.

x402 versus Stripe MPP versus Google AP2 versus OpenAI ACP.
Which rails settle the transaction.

Wrong layer.

Jordan Lyall mapped the whole stack this week. 162 projects, $8.05 billion in acquisitions in 18 months. Stripe bought Bridge for $1.1B, bought Privy.  Mastercard bought BVNK for $1.8B. Capital One bought Brex for $5.15B.

The most contested layer is L4. The governance layer. With a label I'd like you to sit with: "May this agent spend?"

Four serious teams.
Four completely incompatible answers.

World: trust flows from biometric humanness.
Coinbase and ERC-8004: trust flows from on-chain registration.
Skyfire: trust flows from a licensed intermediary. Classic financial services logic, but for robots.
Mastercard and Google: trust flows from cryptographic proof. The audit trail.

a16z's Noah Levine put the honest number out this month: real agent payment volume, after filtering wash trades, is $1.6M per month. $8 billion deployed to capture that.

The bet is not on today. The bet is on who controls the gate when volume arrives.

Nobody has agreed what the gate is supposed to check.

Everyone is fighting about the rails. Nobody has answered the harder question: where does trust come from when the buyer is not human?

CAPTCHAs were annoying and everyone hated them. But they had a clear answer to a simple question.

The agent payments stack has 162 projects, $8 billion, and four incompatible definitions of the thing the gate is supposed to verify.

I have a slide deck titled "Agentic Commerce Readiness." The L4 governance slide is a question mark in 48-point font. I made it bold. I felt that was important.