AgentRails

Anonymous. Tracking how AI agents pay: x402, AP2, ACP, MPP, stablecoin rails. Real volume vs. hype. No sponsors.

[AgentRails @AgentRails_]

Every big payments company now has an "AI agent payments" product.
Visa. Mastercard. Stripe. Google. Coinbase. OpenAI.
Almost none of them tell you how much real money is actually moving.
This account does one thing: tracks what's real in agent payments, and what's still a slide deck.
Weekly numbers. Plain English. No sponsors.

[AgentRails @AgentRails_]

@lukemulks @brave @BraveSearchAPI Search is the most natural paid call for an agent: small, frequent and clearly worth it. Supporting both x402 and MPP instead of picking a side is the smart move too. Would love to see what share of paid queries ends up coming from agents vs humans once it's live.

[AgentRails @AgentRails_]

@roasbeef Per-token metering is the part most pay-per-call setups skip. Pay per request means you prepay for an answer you haven't seen. Streaming means you stop paying the second the model goes off the rails. Curious how the session accounts for a stream that gets cut mid-answer.

[AgentRails @AgentRails_]

@alaxic Sharp observation. The catch: to an issuer's fraud model, "decline, then instantly try three more cards" is the classic card-testing pattern. Without a way to tell a legit agent retry from a fraud bot, merchants may start blocking the exact agents trying to complete the sale.

[AgentRails @AgentRails_]

@agentwormhole @RaetoCC 7 stopped out of 43 is a 16% flag rate, which is wild for payee checks. Curious what those 7 had in common: brand-new wallets, known bad addresses, or metadata that didn't match the service? That breakdown might be the most useful number in agent security right now.

[AgentRails @AgentRails_]

@Tollxyz_ Respect for stripping out the router loop instead of leading with 25K. Same pattern on Base: one service did a third of all x402 calls for about 1.8% of the dollars. Your 27-cent average is in the same range as the low-30s we see there. Bookmarking the Monday refresh.

[AgentRails @AgentRails_]

@CognitionWire @NashBriggsy Agreed. "Loop until proven otherwise" should be the default for any x402 stat. On Monday I'll put payments-per-buyer next to every service so the loops are easy to spot.

[AgentRails @AgentRails_]

4.72M agent payments last month. Total: $1.75M.

The No. 1 service handled 1 in 3 but just 1.8% of the dollars. About 2 cents each.
No. 2: 620K payments from 161 buyers.

Counts measure activity. Dollars measure demand. x402 has a lot more of the first. (x402scan, 30d)

[AgentRails @AgentRails_]

Exactly the right question. As far as I can see x402scan doesn't split new vs repeat wallets out of the box, but payments per buyer for each service is a decent proxy. I'll break it down in Monday's scoreboard: which services have real repeat buyers, and which are a handful of wallets on a loop.

[AgentRails @AgentRails_]

Source: https://t.co/92YeIzT34b (30-day view). Happy to be proven wrong. Show me the dollars.

[AgentRails @AgentRails_]

@AgentsThatWrkAI Good catch, thanks. Coinbase's own tweet lumped product permissions in with the live controls, but the post is the source of truth. And server-side is the key part: a limit the agent can't edit is the only kind that counts. Hoping permissions ship soon.

[AgentRails @AgentRails_]

Coinbase just gave trading agents guardrails: max trade size, a 24h volume cap, auto-stop, product permissions.

The 24h cap is the one that matters. A per-trade limit doesn't stop a loop. An agent can make 300 small trades under it.

Still missing: a per-session budget.

[AgentRails @AgentRails_]

Agreed, and that’s exactly where agents break it. The IP and device are the agent’s server, not the user’s phone. Every timestamp lines up perfectly with a bot. The only signal that still ties back to the person is the approval itself, which is why signed mandates like AP2’s matter.

[AgentRails @AgentRails_]

Agent checkout has a dispute gap.

A user this week said her AI assistant booked an appointment and started a checkout before she approved.

You click buy, the card network knows you meant it. An agent clicks, the merchant can't prove you did. Someone eats that chargeback.

[AgentRails @AgentRails_]

@RWAFoundation_ @aave @sparkfinance @Morpho @Uniswap @VenusProtocol @0xfluid @pendle_fi @tokenterminal Interesting split: total stablecoin supply has been shrinking for weeks, yet lending TVL keeps climbing. Dollars aren't leaving so much as moving from trading to yield. Parked dollars earn, but they don't pay for anything. Would love the same chart for payment volume.

[AgentRails @AgentRails_]

@KasperoLabs 23,874 hostile transactions is the kind of number more teams should publish. Curious about the x402 piece: when an agent pays with any token, who eats the slippage if the swap moves between quote and settle? That's where most pay-with-anything designs get messy.

[AgentRails @AgentRails_]

@its_veteran1 @IPXExchange Love that you're framing it as a claim to test. You're exactly the person who can test it. Would genuinely like to see your numbers after a month: average days to get paid and total fees, old way vs Bridge on Base. The cash-out to naira is usually where the savings leak.

[AgentRails @AgentRails_]

@chargeflow Breaking disputes down by reason is going to get interesting fast. Feature request for 2027: a filter for orders placed by an AI agent. "I didn't authorize that, my assistant did" is about to become its own dispute category, and merchants will want to see the win rate on it.

[AgentRails @AgentRails_]

@jshah321 This matters more for agents than for traders. A company letting its agents pay in USDC needs it on the balance sheet as cash, not as a financial asset it has to mark and test. Get the classification wrong and every CFO quietly says no to agent wallets.

[AgentRails @AgentRails_]

Jensen's job search, but make it agentic:

His agent finds the offer, negotiates the package, signs with a verified ID and pays the moving truck in USDC.

He just has to show up at Starbase.

Agent payments, finally useful.

[AgentRails @AgentRails_]

@gizmocloud_ark Solid list. The one most teams skip is the audit trail. Limits stop the damage, but the log is what lets you prove what the agent actually did when a user or merchant disputes it. And it matters who signs that log, not just who stores it.

[AgentRails @AgentRails_]

@moonpay @xpticket @PayBox Concert tickets are the final exam for agent commerce: surge pricing, scalper bots and zero refunds. If an agent can land two seats without paying a reseller triple, it can buy anything. If it can't, "make no mistakes" is going to do a lot of heavy lifting.

[AgentRails @AgentRails_]

@HalseyHuth @moonpay @PayBox Tiger teams are the right instinct. First question I'd hand every one of them: what share of your checkouts already come from agents? Most merchants can't tell yet, and it's hard to build for traffic you can't even see. "FDA" is elite, by the way.