Agent Pay Opens Its Wallet—AI Agents Still Mostly Spend Pennies

Mastercard opened early access to Agent Pay for Machines, a system that lets autonomous software discover services, agree to terms and pay for them within limits set by a human, after unveiling a live transaction at Cloudflare’s Connect conference. The company first announced the system in June, and Mastercard’s Jorn Lambert demonstrated the flow. Mastercard is also releasing buyer- and seller-side APIs, a toolkit and a sandbox for developers.

Agent Pay for Machines builds on x402, the open protocol that lets online services request payment, and pairs it with Mastercard’s network. The design addresses AI agents becoming “economic actors” that need to buy data from one provider, an AI model from another and computing power from a third without opening conventional accounts or billing relationships. A business gives an agent a task and spending cap; when a service returns a “402 Payment Required” response, the agent generates a digitally signed voucher, a secure promise to pay within its approved limit, which the provider verifies before delivering the service. The interactions happen off-chain at machine speed, and providers can batch vouchers to redeem later.

In the demonstration, an agent for marketing agency Hype planned a protein-bar campaign on a set budget, discovered registered providers and executed a two-step plan spanning social media research from one vendor and campaign images via Cloudflare. The run covered nine payments to two independent merchants with no separate accounts or invoices. Coinbase’s developer platform supplied the wallet, the Ethereum layer-2 network Base provided on-chain settlement, and BVNK, the stablecoin firm Mastercard acquired in August, handled payment and settlement in fiat or stablecoins through Mastercard’s Global Money Platform. The launch follows end-to-end pilots with Cloudflare, Ant International, AWS, BVNK and Coinbase, as well as Mastercard’s recent launches of the Stablecoin Security risk tool and the BVNK-powered Global Money Platform.

A BeInCrypto Intelligence report published in collaboration with AWS, Cloudflare, Chainalysis and Zoomex found that actual AI-agent payment activity remained heavily weighted toward micropayments. The State of AI Agent Payments 2026 tracked payments and AI financial tools across a 35-day study of 6.4 million transactions, in which $119,947 changed hands. Ninety-point eight percent of sampled x402 payments were below one cent, with much of the activity involving API access, computing resources and automated data requests. On the report’s 100-point assessment, micropayments scored 60.7, while institutional financial workflows scored 12.0. Humans still controlled most financial decisions: 14 of 19 reviewed AI tools operated within preset spending or trading limits, AI-labeled interfaces represented less than 0.55% of Hyperliquid’s trading volume, Zoomex order-preparation workflows required human confirmation, and Six of 12 institutional workflows could initiate or settle payments within authorized limits.

BeInCrypto asked three industry leaders what stands between micropayments and serious capital, including THORWallet founder and CEO Marcel Robert Harmann and ZeruAI founder and CEO Ashutosh Sahoo. Harmann expects more agentic trading in the next 12 months as venues such as Hyperliquid offer APIs for automated execution, but said controlling funds remains the central challenge. “Microtransactions, where agents pay for data or compute, will continue to dominate the transaction count. Volume will likely increase as more actual trading runs through agents.” He proposed threshold signatures as a way to limit unrestricted wallet access. “With threshold signatures, the agent can hold one key share and an independent policy signer can hold a second one. We can adapt our workflows to the new situation.” Harmann also expects regulators to adopt automated monitoring as transaction counts rise, arguing that humans cannot examine every machine payment individually and oversight systems must identify suspicious activity at the same pace while people retain final authority.