An AI agent made nine payments to two service providers within a single approved budget in a demonstration of Mastercard’s Agent Pay for Machines at Cloudflare Connect. On October 20, 2026, Mastercard opened early access to the system, which lets agents buy digital services for businesses while keeping spending and permissions under human control.
Key takeaways
- Signed vouchers let agents spend within business-approved limits.
- Hype’s demonstration purchased research and eight campaign images.
- Early Access gives developers payment APIs and testing tools.
The developer opening follows end-to-end pilots involving Cloudflare, Ant International, AIsa, Base, BVNK, Coinbase and Hype. The project remains at the pilot stage rather than being commercially available.
Mastercard frames the central challenge as authority: establishing which agent can spend, on what services and within which limits. Its system connects the open x402 payment protocol with the Mastercard network’s trust and security infrastructure.
How Mastercard’s Agent Pay for Machines authorizes spending
Agent Pay for Machines uses digitally signed payment vouchers to authorize purchases without giving an AI agent unrestricted spending power. A business first defines the task, permissions and budget.
When the agent reaches a paid service, the provider returns “402 Payment Required”. The agent then signs a voucher promising payment within its authorized limit. The provider checks that voucher before supplying the service.
Mastercard says these exchanges happen off-chain at machine speed. Providers can group vouchers for later redemption instead of settling every purchase individually, while agents avoid establishing a separate billing relationship for each service.
The underlying x402 protocol, incubated by Coinbase, supports online payment requests through blockchain-based and traditional payment rails. Mastercard’s system can support purchases of AI models, data, computing power, APIs and other digital services.
Hype’s pilot bought research and campaign images
Marketing agency Hype demonstrated an agent planning and executing a sample campaign for a protein bar launch, with purchases restricted to an approved task and budget.
The agent discovered registered providers and proposed two steps: buying social media research from AIsa and creating campaign images through Cloudflare AI Gateway. Mastercard described the workflow as running on an agent platform such as Amazon Bedrock AgentCore.
Hype approved the funding before the agent started spending. Payment requests arrived through Cloudflare Monetization Gateway; the agent signed vouchers, and the sellers verified them before continuing the work.
The demonstration produced research followed by eight campaign images. In total, the agent made nine payments across two independent merchants, without separate accounts, API keys or invoices, Mastercard said.
Coinbase, Base and BVNK supply the payment infrastructure
Coinbase Developer Platform supplied the wallet infrastructure for funding the agent, while Base provided on-chain settlement. BVNK served as the agent wallet.
Payments and settlement for both parties, whether in fiat or stablecoins, were handled by BVNK through the Mastercard Global Money Platform. This setup allows sellers to receive funds in whatever currency, token or payment rail they prefer.
The arrangement separates the agent’s off-chain voucher exchanges from the settlement infrastructure behind them. Mastercard connects agent platforms, wallets, payment providers and merchants through a shared trust layer.
Early Access opens developer testing
Agent Pay for Machines Early Access provides the first buyer- and seller-side APIs, alongside a toolkit, sandbox environments and developer documentation.
Mastercard is asking developers to test agent-to-agent commerce use cases and provide feedback on features they need. It says customer and partner requirements will guide the roadmap, with an open, interoperable and trusted agent economy as its stated goal.
Businesses using the system define the agent’s assignment, permissions and spending ceiling; the agent is authorized to transact only within those parameters.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.