Amazon Bedrock AgentCore payments 正式发布：让 Agent 能够安全、自主且大规模地完成交易

Agent 已从简单的聊天应用演变为可长时间运行的自主系统：它们能够在无人监督的情况下，为每项任务动态发现并组合数十种工具。与此同时，服务和内容提供商正从以人为中心、基于订阅的一刀切定价，转向按使用量、按执行次数计费的模式，单次成本通常仅为几美分。如今，Agent 在推理、选择工具和完成任务方面表现出色，但一旦涉及支付便会遇到障碍。为此，我们于 5 月与 Coinbase 和 Stripe 合作推出了 AgentCore payments 预览版，让 Agent 开发者只需几行代码，即可为其 Agent 配备自主支付付费 API、MCP 和内容的能力。今天，我们很高兴地宣布，该服务现已正式发布，可让企业凭借安全性、防护机制和可观测性，为生产工作负载提供 Agent 支付能力。

主要功能

钱包支持

Agent 需要资金来源才能执行 Agent 支付交易。AgentCore payments 通过集成 Coinbase 和 Stripe Privy 钱包解决这一问题，开发者只需提供 Coinbase 或 Privy 的 API 密钥及密钥材料，即可配置支付能力。这些是专为低成本微支付而设计的稳定币钱包，交易金额通常低至几美分。终端用户可通过信用卡等传统支付方式或 USDC 稳定币为其 Agent 钱包注资，并且必须授权 Agent 代表其进行消费。

为验证安全性，AgentCore 会将开发者凭证存储在 AgentCore Identity Secrets Manager 中。Agent 无法看到原始凭证。AgentCore payments 使用由这些凭证派生的短期令牌，按照开发者 Agent 的指令要求钱包提供商执行钱包操作（例如交易签名），从而为交互增加额外的安全层。

针对 GA 版本，我们直接在 AgentCore Payments 中构建了适用于 Coinbase 的“Quick Create”选项（可通过控制台或 CLI 使用），开发者无需离开 AgentCore 即可配置凭证，从而节省时间和精力。对于 Stripe Privy，开发者需要从 Privy 控制台获取凭证，并将其提供给 AgentCore payments。

支付编排

AgentCore payments 在设计上与协议无关，通过抽象底层复杂性，让开发者只需集成一次，即可与采用不同协议的商家协作。在预览版阶段，我们推出了对 x402 协议的支持。在 GA 版本中，我们将协议支持扩展至 Machine Payment Protocol（MPP）——这是另一项机器支付标准，由 Stripe 和 Tempo 共同编写——开发者无需额外增加一行代码，即可为任何兼容 MPP 的服务付款。

此外，我们还在 x402 中引入了对“upto”方案的支持，使 Agent 能够设置消费上限，而非承诺固定价格。这解锁了真正的按推理付费和动态定价：提供 LLM token、计算能力或任何按用量计费 API 的商家，现在可以在调用结束时根据实际消费量准确收费，而不再需要像当前支持的“exact”方案那样为请求设定固定价格。

按使用量端点的可发现性

AgentCore 通过 AgentCore Gateway，将 Coinbase 提供的按使用量计费 x402 端点以 MCP 服务器形式暴露，使 Agent 开发者能够根据自身用例发现并选择相关的付费端点。现在，我们通过将 MCP 服务器更新为经过筛选的高质量端点列表，进一步增强了这一发现体验；筛选依据包括社会认同、元数据丰富度、描述质量和可用性。

支付限额

Agent 天生具有非确定性，因此可能将某个响应误解为消费授权，或因意外重试而重复付款。在 AgentCore payments 中，交易在支付会话内运行；支付会话是针对单次 Agent 交互的、具有范围限制的支付上下文。一个支付会话具有两项可配置上限：指定货币的最高消费金额，以及过期时间。在签署支付前，AgentCore payments 会根据会话预算检查请求，并拒绝会使会话超过限额的请求。该检查具有确定性，并在基础设施层执行。

可观测性

当 Agent 执行自主交易时，开发者需要全面的可观测性——包括支付审计轨迹、详细日志，以及成功率和平均交易金额等关键指标——以保持对 Agent 消费行为的可见性并进行审计。AgentCore payments 与 AgentCore Observability 集成，使开发者能够查看支付生命周期；同时，它会自动将托管日志发送至您的 Amazon CloudWatch 日志组，并将托管 span 发送至 AgentCore Observability。AgentCore Observability 包含预构建仪表盘，可跨 Agent、支付会话和时间周期展示端到端交易健康状况（例如交易成功率、平均交易金额）。

使用场景

支付付费网页内容

如今，在 Web 上托管内容的出版商正在阻止 Agent 访问，并将 Agent 流量货币化。无论是用于信息检索还是基于 Web 的自动化，部署实时访问 Web 的 AI Agent 的企业、AI 浏览器工具提供商以及 Agent Web 搜索提供商，都将受益于集成 AgentCore payments 服务；该服务使其 Agent 解决方案能够在不中断 Agent 工作流的情况下，为付费墙后的网页内容付费。为实现这一目标，我们持续与包括 Amazon CloudFront 和 Cloudflare 在内的内容分发网络（CDN）合作。我们的客户之一 Anchor Browser 已集成 AgentCore payments，为其客户解锁付费墙后的网页内容。

“Anchor Browser 是面向企业的领先云端浏览器自动化平台之一，每日处理数百万流量，专为 AI Agent 构建。我们很高兴集成 AgentCore payments，使客户能够在基于 Web 的 Agent 工作流中解锁付费内容。集成过程很顺利，并且它提供了一种安全的方式来处理 Agent 交易，同时内置支付限额。”

— Anchor Browser 联合创始人兼 CEO Idan Raman

“Cloudflare 位于全球相当大一部分内容、API 和 MCP 服务器的前方。我们的 Monetization Gateway 让这些资产的所有者可以使用开放的 x402 协议，按请求并按其设定的价格直接向 Agent 收费。借助 AgentCore payments，在 AgentCore 上构建 Agent 的企业从第一天起即可访问并支付这些资产。”

— Cloudflare 首席战略官 Stephanie Cohen

按推理付费

按推理付费是 Agent 交易的新兴用例，使 Agent 开发者能够在不同模型之间动态切换，以在不牺牲质量的前提下优化成本，同时避免管理各个模型提供商 API 密钥和订阅所带来的额外负担。我们的客户之一 SpreadX 已在其新 AI 产品 Incarna 中使用 AgentCore Payments，通过 BlockRun 支付推理费用；BlockRun 是一个路由和支付层，Agent 可通过 x402 协议按调用付费。

“我们很高兴与 AgentCore payments 合作。该托管服务让 Agent 能够轻松、安全地访问并支付 BlockRun 的按使用量付费端点，从而解锁全新的按推理付费用例。”

— BlockRun 创始人兼 CEO Vicky Fu

消费者与深度研究用例

旅游服务商 Travala 已将 AgentCore payments 集成至其 Travala Travel MCP 服务器，使使用 Claude 等 Agent 平台的终端客户能够在单次对话中安全、可靠地以对话方式预订酒店。类似地，Elsa AI 和 Heurist AI 通过 x402 协议访问按使用量付费 API，借助 AgentCore payments 向终端客户提供金融研究和咨询服务。

“将 AgentCore payments 集成到 Travala Travel MCP 中，简化了我们全球 220 万处房源的 Agent 预订流程。该集成提供了一种安全的方式来处理 Agent 交易，并内置支付限额，从而改善了客户整体的 Agent 预订体验。”

— Travala CEO Juan Otero

开始使用 AgentCore payments

AgentCore payments 已在以下区域推出。开发者体验与选择权是 AgentCore payments 的基础设计原则。通过即用型技能、直观的控制台以及与流行开源框架的集成，我们让开发者能够轻松使用 AgentCore payments 构建应用，从而专注于最重要的事情：安全且大规模地提升其采购 Agent 的质量和性能。 您可以通过以下多种方式设置和使用 AgentCore payments：

编码助手技能

无论您是首次构建可交易 Agent，还是为现有 Agent 添加自主支付功能，都可以让您的编码助手（Claude Code、Kiro、Codex）使用该技能。您的编码助手将获得构建支持支付 Agent 所需的完整上下文——从凭证设置到交易执行——使您的 Agent 开箱即可自主支付付费服务或内容。更多详情请参阅快速入门指南。

AgentCore CLI

安装 AgentCore CLI，只需几条命令即可构建脚手架、配置并部署具备完整交易能力的 Agent。它会连接支付能力——包括凭证配置和交易路由——让您获得一个端到端、支持支付且可在数分钟而非数天内完成交易的 Agent。更多详情请参阅快速入门指南。

AWS Management Console

您可以通过 Amazon Bedrock AgentCore 控制台开始使用。在导航窗格中的 Build 下，选择 Payments。Quick Create 选项可让您在控制台中创建 Coinbase 凭证并设置 AgentCore payments 资源。

开源框架集成

构建可交易 Agent 时，开发者需要在其 Agent 框架的执行循环中处理被拦截的 HTTP 402 Payment Required 响应。AgentCore payments 可与流行的开源框架及任何模型配合使用。AgentCore payments 提供的集成只需极少代码修改，即可添加支付能力。请在以下部分探索特定框架的实现：

plugin = AgentCorePaymentsPlugin(config=AgentCorePaymentsPluginConfig(
    payment_manager_arn=os.environ["PAYMENT_MANAGER_ARN"],
    user_id="test-user-123",
    payment_instrument_id=os.environ["PAYMENT_INSTRUMENT_ID"],
    payment_session_id=os.environ["PAYMENT_SESSION_ID"],
))
agent = Agent(model=BedrockModel(model_id=MODEL_ID, streaming=True), tools=[], plugins=[payment_plugin], system_prompt=SYSTEM_PROMPT)

payments = AgentCorePaymentsMiddleware(config)
agent = create_agent(model, tools=[], system_prompt=SYSTEM_PROMPT, middleware=[payments])

agents-pay 技能使 Agent 能够在运行时为受 x402 保护的内容付款——通过 AgentCore payments 在任务进行期间拦截 402 Payment Required 付费墙，使任务无需人工干预即可自主完成。

结论

今天，AgentCore payments 正式发布，这是让 Agent 经济成为现实的一步。在这种经济中，AI Agent 不仅能够推理和行动，还能够以自主、安全且可扩展的方式完成交易。通过抽象支付协议、钱包管理和交易编排的复杂性，我们使开发者能够专注于构建智能 Agent，而由 AgentCore 处理基础设施。我们的客户已经证明：当支付成为 Agent 技术栈原生的一等能力时，其可能性具有广泛的覆盖范围。凭借与协议无关的支持、内置消费防护机制和生产就绪的可观测性，AgentCore payments 让开发者有信心在数分钟而非数天内将可交易 Agent 部署到生产环境。立即开始使用，为您的 Agent 配备交易能力。

资源

---

关于作者

Chethan Shriyan

Chethan 是 AWS 的首席技术产品经理，拥有超过 12 年的产品和业务管理经验。Chethan 热衷于构建和交付能够为客户生活带来切实影响的技术产品。

Madhu Samhitha Vangara

Madhu 是 AWS 全球生成式 AI 专项解决方案架构师，专注于 Amazon Bedrock AgentCore 和 Strands Agents 的 Agentic AI 技术 GTM。她拥有深厚的企业经验，擅长将新兴 AI 能力转化为可衡量的客户成果。Madhu 是 AI 会议演讲者，专长于生产级 Agentic AI。

Agents have evolved from simple chat applications to autonomous, long-running systems that dynamically discover and compose dozens of tools per task without human oversight. On the other side, service and content providers are moving from human-centric subscription-based, one-size-fits-all pricing to pay-per-use, per-execution models where costs are often a few cents. Today, agents are doing a great job at reasoning, choosing tools, and completing tasks, but they hit a roadblock when payment is involved. To address this, we launched AgentCore payments in collaboration with Coinbase and Stripe in preview in May, enabling agent developers to equip their agents to autonomously pay for paid APIs, MCPs, and content with a few lines of code. Today, we are happy to announce that this service is generally available, enabling enterprises to power agentic payments with security, guardrails, and observability for production workloads.

Key features

Wallet support

Agents need a source of funds to make agentic transactions. AgentCore payments addresses this by integrating with Coinbase and Stripe Privy wallets, enabling developers to set up payment capabilities by providing Coinbase or Privy API keys and secrets. These are stablecoin wallets purpose-built for cost-effective microtransaction payments, often in cents. End users can fund their agent’s wallet through traditional payment methods like credit cards or through USDC stablecoin and must grant delegation to the agent to spend on their behalf. To verify security, AgentCore stores developer credentials in the AgentCore Identity Secrets Manager. The agent does not see the raw credentials. AgentCore payments uses short-lived tokens derived from these credentials to instruct the wallet provider to perform wallet operations (such as transaction signing) as directed by the developer’s agent, adding an extra layer of security for the interactions.

For GA, we built a “Quick Create” option for Coinbase directly within AgentCore Payments (available through the console or CLI) that enables developers to provision credentials without leaving AgentCore, saving developers time and effort. For Stripe Privy, developers need to obtain the credentials from the Privy dashboard and provide those to AgentCore payments.

Payment orchestration

AgentCore payments is built to be protocol-agnostic, abstracting away the complexity so developers can integrate once and work with merchants regardless of which protocol they use. At preview, we launched with support for the x402 protocol. At GA, we have expanded protocol support to include the Machine Payment Protocol (MPP), another standard for machine payments co-authored by Stripe and Tempo, enabling developers to pay for any MPP-compatible service without an additional line of code. Additionally, we have introduced support for the “upto” scheme within x402, which enables an agent to set a spending ceiling rather than committing to a fixed price. This unlocks true pay-per-inference and dynamic pricing: a merchant serving LLM tokens, compute, or any usage-metered API can now charge for exactly what was consumed at the end of a call, rather than requiring a fixed price for requests (as with the “exact” scheme supported today).

Discoverability of pay per use endpoints

AgentCore exposes pay-per-use x402 endpoints offered by Coinbase as an MCP server via AgentCore Gateway, enabling agent developers to discover and select relevant paid endpoints based on their use case. We have now enhanced this discovery experience by updating the MCP server to a curated list of high-quality endpoints based on social proof, metadata richness, description quality, and availability.

Payment limits

Agents are inherently non-deterministic, so they can misinterpret a response as authorization to spend or repeat a payment because of an unexpected retry. In AgentCore payments, transactions run within a payment session, a scoped payment context for a single agent interaction. A payment session has two configurable caps: a maximum spend amount in a specified currency, and an expiry time. Before signing a payment, AgentCore payments checks the request against the session budget and rejects requests that would push the session past its cap. The check is deterministic and runs at the infrastructure layer.

Observability

As agents execute autonomous transactions, developers need comprehensive observability—including payment audit trails, detailed logs, and key metrics such as success rates and average transaction values—to maintain visibility and audit their agents’ spending behavior. AgentCore payments integrates with AgentCore Observability to give developers visibility into the payment lifecycle and automatically emits vended logs to your Amazon CloudWatch log group, and vended spans to AgentCore Observability. AgentCore Observability includes prebuilt dashboards that show end-to-end transaction health (for example, transaction success rate, average transaction value) across agents, payment sessions, and time periods.

Use cases

Pay for paid web content

Publishers hosting content on the web today are blocking agent access and monetizing agent traffic. Enterprises deploying AI agents that access the web in real time, whether for information retrieval or web-based automation, AI browser tool providers, and agentic web search providers would benefit from integrating with the AgentCore payments service, which enables their agentic solutions to pay for paywalled web content without breaking the agent workflow. To enable this, we continue to collaborate with Content Delivery Networks (CDN) including Amazon CloudFront and Cloudflare. One of our customers, Anchor Browser, integrated with AgentCore payments to unlock paywalled web content for their customers.

“Anchor Browser is one of the leading cloud-based browser automation platforms for enterprises with millions of traffic per day, purpose-built for AI agents. We are excited to integrate AgentCore payments to enable our customers to unlock paid web content within their web-based agentic workflows. The integration was smooth, and it offers a secure way to handle agentic transactions with built-in payment limits.”

— Idan Raman, CEO, Co-Founder of Anchor Browser

“Cloudflare sits in front of a significant share of the world’s content, APIs, and MCP servers. Our Monetization Gateway lets owners of those assets charge agents directly, per request and at a price they set using the open x402 protocol. With AgentCore payments, enterprises building agents on AgentCore can access and pay for these assets on day one.”

— Stephanie Cohen, Chief Strategy Officer, Cloudflare

Pay per inference

Pay-per-inference is an emerging use case for agentic transactions, enabling agent developers to dynamically switch between models to optimize for cost without compromising on quality and the overhead of managing API keys and subscriptions with individual model providers. One of our customers, SpreadX, used AgentCore Payments in their new AI product, Incarna, to pay for inference through BlockRun — a routing and payment layer where agents pay per call over the x402 protocol.

“We’re thrilled to collaborate with AgentCore payments, which offers a managed service that makes it easy for agents to securely access and pay for BlockRun’s pay-per-use endpoints, unlocking a whole new pay-per-inference use case.”

— Vicky Fu, Founder and CEO, BlockRun

Consumer and deep research use case

Travala, a travel provider, has integrated AgentCore payments into their Travala Travel MCP servers, enabling end customers using agentic platforms like Claude to safely and securely book hotels conversationally in a single chat. Similarly, Elsa AI and Heurist AI use AgentCore payments to deliver financial research and advisory to their end customers by accessing pay-per-use APIs over the x402 protocol.

“Integrating AgentCore payments into the Travala Travel MCP streamlines the agentic booking process for our 2.2 million properties globally. This integration offers a secure way to handle agentic transactions with built-in payment limits, improving the overall agentic booking experience for our customers.”

— Juan Otero, CEO of Travala

Get started with AgentCore payments

AgentCore payments is available in the following regions. Developer experience and choice is the foundational design principle of AgentCore payments. With ready-to-use skills, an intuitive console, and integrations with popular open-source frameworks, we make it straightforward for developers to build with AgentCore payments so they can stay focused on what matters most: the quality and performance of their buyer agents securely at scale. You can set up and use AgentCore payments through multiple options:

Coding assistant skill

Whether you are building a transacting agent for the first time or adding autonomous payment capabilities to an existing agent, point your coding assistant (Claude Code, Kiro, Codex) to the skill. Your coding assistant will have the full context it needs to scaffold payment-enabled agents — from credential setup to transaction execution — so your agent can autonomously pay for paid service or content out of the box. For more details, refer to the Quick start guide.

AgentCore CLI

Install the AgentCore CLI to scaffold, configure, and deploy fully transacting agents in a few commands. This wires up payment capabilities — including credential provisioning, transaction routing — giving you an end-to-end payment-enabled agent ready to transact in minutes, not days. For more details, refer to the Quick start guide.

AWS Management Console

You can get started with the Amazon Bedrock AgentCore console. In the navigation pane, under Build , choose Payments. The Quick Create option provides a way to create your Coinbase credentials on the console and set up your AgentCore payments resources.

Open source framework integrations

When building agents that transact, developers need to handle the HTTP 402 Payment Required response intercepted within their agent framework’s execution loop. AgentCore payments works with popular open source frameworks and any model. AgentCore payments provides integrations that add payment capabilities with minimal code changes. Explore the framework-specific implementations in the following section:

plugin = AgentCorePaymentsPlugin(config=AgentCorePaymentsPluginConfig(
    payment_manager_arn=os.environ["PAYMENT_MANAGER_ARN"],
    user_id="test-user-123",
    payment_instrument_id=os.environ["PAYMENT_INSTRUMENT_ID"],
    payment_session_id=os.environ["PAYMENT_SESSION_ID"],
))
agent = Agent(model=BedrockModel(model_id=MODEL_ID, streaming=True), tools=[], plugins=[payment_plugin], system_prompt=SYSTEM_PROMPT)


payments = AgentCorePaymentsMiddleware(config)
agent = create_agent(model, tools=[], system_prompt=SYSTEM_PROMPT, middleware=[payments])

The agents-pay skill enables an agent to pay for x402-protected content at runtime — intercepting a 402 Payment Required paywall mid-task via AgentCore payments, allowing autonomous task completion without human intervention.

Conclusion

Today, AgentCore payments is generally available, marking a step in making the agentic economy a reality, where AI agents don’t just reason and act, but transact autonomously, securely, and at scale. By abstracting the complexity of payment protocols, wallet management, and transaction orchestration, we enable developers to focus on building intelligent agents while AgentCore handles the infrastructure. Our customers are already proving the breadth of what’s possible when payments become a native, first-class capability of the agent stack. With protocol-agnostic support, built-in spending guardrails, and production-ready observability, AgentCore payments gives developers the confidence to deploy transacting agents in production in minutes, not days. Get started today and equip your agents to transact.

Resources

---

About the authors

Chethan Shriyan

Chethan is a Principal Product Manager – Technical at AWS. He has 12+ years of experience in product and business management. Chethan is passionate about building and delivering technology products that create meaningful impact in customers’ lives.

Madhu Samhitha Vangara

Madhu is a Worldwide generative AI Specialist Solutions Architect at AWS, focusing on Agentic AI technical GTM for Amazon Bedrock AgentCore and Strands Agents. She brings deep enterprise experience translating emerging AI capabilities into measurable customer outcomes. Madhu is a speaker at AI conferences and specializes in production-grade Agentic AI.