回顾：2026 年 8 月面向 AI 开发者的 Amazon Bedrock、AgentCore 与 Strands 更新
英文标题：ICYMI: What landed for AI builders in August 2026
AWS 回顾了 2026 年 8 月 Amazon Bedrock、Amazon Bedrock AgentCore 和 Strands 面向 AI 开发者推出的主要能力，包括 OpenAI 模型的百万 Token 上下文、联网搜索和跨区域推理。AgentCore 新增最长可持续 14 天的专用计算运行时、时序策略、支付控制、网页搜索、记忆隔离与 Agent Registry，以支持更长时间、更可控的生产级 Agent 工作流。AWS GovCloud（US）现可使用 Claude Opus 5、OpenAI GPT-5.6 Terra 和 Luna，以及多模态检索和 AgentCore 能力，为受监管工作负载扩展了选择。Strands Robots 则将 Agent、机器人数据采集、策略训练和实体部署连接到统一工作流，推动多机器人及物理 AI 应用开发。
1.3 MB
101.1 KB
124.1 KB
345.0 KB
141.2 KB
346.5 KB
18.8 KB
126.7 KB
中文
回顾 2026 年 8 月 Amazon Bedrock、AgentCore 和 Strands 的重要更新。
在 AWS，我们长期致力于让基础技术变得易于获取，并提供将其投入实际应用所需的基础设施。已有超过 22.5 万活跃客户使用的 Amazon Bedrock——其中包括逾 80% 的《财富》100 强企业——正体现了这一承诺：它提供领先模型及广泛的工具集，帮助客户构建和扩展 AI，同时满足生产环境所需的安全性、可靠性、性能和成本效率。随着 Agent 应用不断扩展，Amazon Bedrock 通过 AgentCore 进一步夯实这一基础；它允许您使用任意框架和模型构建、连接及优化 Agent。AWS 还将 Strands Agent Harness SDK 作为开源项目发布，让您能够灵活创建 Agent，并将其部署到您选择的任何位置。它们共同帮助您将模型智能转化为生产级 Agent，而不限制这些 Agent 的构建方式或运行地点。
随着模型能力不断增强，组织可以提出一个更具雄心的问题：AI 从头到尾究竟能够承担多少有实际意义的工作？跨行业来看，Agent 正进入这样的工作流：它们需要基于企业信息进行推理，在更长的时间跨度内保持高效，并跨越数字与物理系统执行操作。随着这类工作范围扩大，关注点正转向模型周围的完整系统，包括它能够访问哪些上下文、可以采取哪些行动、数据在哪里处理，以及如何治理其决策和成本。
我们相信，AI 的下一阶段不仅取决于模型能做什么，还取决于您能以多大的信心将责任委托给围绕模型构建的系统。这种信心依赖于能够理解当前任务、在任务持续推进时保持运行，并在明确的组织边界内运作的系统。8 月的更新正朝着这一方向迈进，为 AI 系统提供更坚实的基础，使其能够理解更多信息、工作更久，并在企业、受监管和物理环境中承担高要求的责任。
以更强控制力大规模推理更多上下文
**在 Amazon Bedrock 上为您的 OpenAI 应用引入完整工作集和最新网页上下文。**GPT-5.6 Sol、Terra 和 Luna 现已支持 百万 Token 上下文窗口，并支持提示词缓存，可在重复使用上下文时帮助降低成本和延迟。Web Search 让模型能够查找其训练数据之外的最新信息，将相关结果纳入回答，并提供引用，无需您集成独立的搜索服务提供商。对于需要最新细节的使用场景，例如实时定价或新发布的文档，这些模型还可以直接检索内容，来源为公开网站。这些能力结合后，您可以分析整个代码库或监管文档，将其内容与最新公开信息进行核对，并通过一次 API 调用返回带引文的回答。
在全球路由推理**，无需自行拼接各个 Region。**借助跨 Region 推理，您可以从超过 25 个 AWS Region 访问 GPT-5.6 Sol、Terra 和 Luna。全球配置文件以更低的每 Token 价格提供最广泛的容量，而地理配置文件则将推理处理限制在定义好的地理区域内。借助这些更新，您可以在需求高峰期间提升吞吐量，同时选择符合应用数据处理要求的地理边界。
随着模型使用量增长，理解并控制推理支出。IAM 主体成本分摊 可帮助您将推理支出归属到用户、团队、项目、应用程序或成本中心；而 AWS Cost Anomaly Detection 现在可监控 Amazon Bedrock 上第三方基础模型的支出，并在成本发生意外变化时提供根本原因明细。OpenAI 还宣布下调 Amazon Bedrock 上 GPT-5.6 系列模型的价格，其中包括 Luna、Terra 和 Sol。这些能力让您能够更清楚地了解哪些团队正在推动支出，更早发现异常变化，并在 AI 工作负载规模化时支持更明智的决策。
从检测到响应，加快网络防御
**为安全团队在防御性及经授权的进攻性工作流中提供前沿 AI。**OpenAI 的 Daybreak Red 和 Daybreak Blue 现已面向 Amazon Bedrock 上符合资格的客户提供。Daybreak Blue 支持漏洞发现、检测工程和事件响应等防御性工作。Daybreak Red 支持漏洞研究、漏洞利用复现和缓解措施开发等高级授权任务。现在，您可以更快地调查和处理漏洞，同时采用更严格的身份验证、监控、访问控制以及零操作员访问基础设施。
借助始终保持正轨的 Agent 完成更多工作
让生产级 Agent 持续运行足够长时间，以完成持续多日的工作。AgentCore 运行时实例 让您能够在专用 Amazon EC2 计算资源上运行 Agent，包括 GPU 加速、内存优化和计算优化实例，单个会话最长可持续 14 天。AgentCore 还扩展至美国西部（北加利福尼亚）和亚太地区（海得拉巴）。现在，您可以将长期运行的研究、编码和监控工作交给 Agent，为每项工作负载匹配所需的计算资源，并让其在更靠近所服务用户和系统的位置运行。
让 Agent 更独立地行动，同时将其行为和支出控制在既定边界内。时序策略 会根据 Agent 已完成的操作评估每项行动，使您能够强制执行操作序列、前置条件、审批关卡、调用之间的值匹配和数据新鲜度要求。速率限制可按用户或组控制请求数、推理 Token 和并发连接数。AgentCore payments 允许 Agent 访问并支付 API、MCP 资源和付费内容，且具备由基础设施强制执行的支出上限与端到端可观测性。这些更新结合后，您可以赋予 Agent 更大的独立性，同时控制哪些事情可以发生、以何种顺序发生、以什么速率发生，以及在何种预算范围内发生。
**在不削弱数据边界的情况下，为 Agent 提供及时上下文。**AgentCore 中的 Web Search 允许 Agent 纳入或排除特定域名，并按发布日期筛选结果，帮助其从获批准的来源检索最新信息。AgentCore memory 可从活动日志、行为事件、系统事件和其他结构化 JSON 数据中提取长期记忆，而不仅限于对话内容。细粒度访问控制 随后可按用户或租户隔离这些记忆。现在，您可以让 Agent 以可信来源为基础，从业务事件中构建记忆，并保持这些上下文相互隔离，而无需为每个应用程序添加自定义授权逻辑。
查找并复用已获批准的 Agent 和工具，而非重复构建已有能力。AWS Agent Registry 提供了一个可搜索、受治理的目录，其中包含 Agent、MCP 服务器、技能和自定义资源。组织范围内的检测能力可以识别已连接 AWS 账户中 AgentCore Runtime 上的 Agent，以及 AgentCore Gateway 上的 MCP 服务器；获批准的资源还可以显示在 Amazon Quick 中。您可以按意图或名称查找可信能力，在不同工作流中复用它们，并减少 Agent 蔓延和重复开发。
在受监管环境中以更多选择进行构建
将高级模型、Agent 能力和跨模态检索引入受监管工作负载。Claude Opus 5 现已在 AWS GovCloud（US）Region 提供，默认启用零数据保留。OpenAI GPT-5.6 Terra 和 Luna 也已在 AWS GovCloud（US）Region 提供，支持百万 Token 上下文窗口和提示词缓存。Amazon Nova Multimodal Embeddings 将文本、文档、图像、视频和音频的跨模态检索能力带到 AWS GovCloud（US-West）；而 AgentCore memory、policy 和托管 harness 则增加了托管上下文、控制和编排能力。这有助于您在 AWS GovCloud 中，使用更多与商业 Region 相同的能力，构建编码 Agent、文档分析系统、多模态检索应用程序和受治理的 Agent 工作流。
将 AI Agent 延伸至物理世界
通过一个相互连接的工作流，完成从机器人演示到训练和实体部署的转变。Strands Robots 连接了 Strands Agents、LeRobot 和 Hugging Face Storage Buckets，让您能够记录演示、以 LeRobot 格式流式传输数据集、训练策略，并将其部署到模拟或实体硬件，而无需转换底层数据。这使您能够从仿真到实体部署，使用一种数据格式和一个 Agent 工作流来迭代机器人策略。
无需为每个环境重建 Agent，即可协调多个机器人和设备。Strands Robots 支持基于网状网络的发现与协调，覆盖模拟设备和实体设备。Zenoh 可连接同一本地网络上的机器人，而 AWS IoT Core 支持地理上分散的设备群。Strands Robots 和 AWS 也正在参与 Model Hardware Standard 的有限研究预览。这让您可以更轻松地在仿真环境中制作多机器人工作流原型，以更少的代码改动将其迁移到本地和云连接的设备群，并探索面向实体设备的标准化控制方式。
开始使用
探索 Amazon Bedrock，使用 AgentCore CLI 部署 Agent，或通过 Strands Harness SDK 构建您的第一个 Agent。
想了解 Amazon Bedrock 如何支持您的团队？联系我们 以开启交流。
关于作者
English / 原文
A recap of the biggest Amazon Bedrock, AgentCore, and Strands updates from August 2026.
At AWS, we have long focused on making foundational technologies accessible and providing the infrastructure needed to put them to work. Amazon Bedrock, used by more than 225,000 active customers, including over 80% of Fortune 100 companies, embodies that commitment by providing access to leading models and a broad set of tools to build and scale AI, with the security, reliability, performance, and cost efficiency customers need in production. As agentic applications expand, Amazon Bedrock extends this foundation with AgentCore, which lets you build, connect, and optimize agents using any framework and model. AWS also released the Strands Agent Harness SDK as open source, giving you the flexibility to create agents and deploy them wherever you choose. Together, they help you turn model intelligence into production agents without limiting how those agents are built or where they run.
As models become more capable, organizations can ask a more ambitious question: how much meaningful work can AI take on from start to finish? Across industries, agents are moving into workflows that require them to reason over enterprise information, remain productive over longer time horizons, and act across digital and physical systems. As the scope of that work expands, attention is shifting to the complete system around the model, including what context it can access, what actions it can take, where data is processed, and how its decisions and costs are governed.
We believe the next phase of AI will be shaped not only by what models can do, but by how confidently you can delegate responsibility to the systems built around them. That confidence depends on systems that can understand the task in front of them, persist as it unfolds, and operate within clear organizational boundaries. August’s updates move in that direction, giving you stronger foundations for AI systems that can understand more, work longer, and take on demanding responsibility across enterprise, regulated, and physical environments.
Reason over more context at scale with greater control
Bring entire working sets and current web context into your OpenAI applications on Amazon Bedrock. GPT-5.6 Sol, Terra, and Luna now support million-token context windows with prompt caching, helping reduce cost and latency when context is reused. Web Search lets the models find current information beyond their training data, incorporate relevant results into their answers, and provide citations without requiring you to integrate a separate search provider. For use cases that need the freshest details, such as live pricing or newly released documentation, the models can also retrieve content directly from public websites. Together, these capabilities let you analyze an entire codebase or regulatory document, check its contents against the latest public information, and return a cited response from a single API call.
Route inference globallywithout stitching together Regions. With cross-Region inference, you can access GPT-5.6 Sol, Terra, and Luna from more than 25 AWS Regions. Global profiles give you the broadest capacity at lower per-token prices, while Geo profiles keep inference processing within a defined geography. With these updates, you can increase throughput during demand spikes while choosing the geographic boundary that fits your application’s data-processing requirements.
Understand and control inference spend as model usage grows. IAM principal cost allocation helps you attribute inference spend to a user, team, project, application, or cost center, while AWS Cost Anomaly Detection now monitors third-party foundation model spend on Amazon Bedrock and provides root-cause breakdowns when costs change unexpectedly. OpenAI also announced lower pricing across the GPT-5.6 family on Amazon Bedrock, including Luna, Terra, and Sol. This gives you greater visibility into which teams are driving spend, helps you catch unusual changes earlier, and supports more informed decisions as AI workloads scale.
Faster cyber defense from detection to response
Give security teams frontier AI for both defensive and authorized offensive workflows. Daybreak Red and Daybreak Blue from OpenAI are now available to eligible customers on Amazon Bedrock. Daybreak Blue supports defensive work such as vulnerability discovery, detection engineering, and incident response. Daybreak Red supports advanced, authorized tasks such as vulnerability research, exploit reproduction, and mitigation development. You can now investigate and address vulnerabilities faster while applying stronger identity verification, monitoring, access controls, and zero-operator-access infrastructure.
Get more done with agents that stay on track
Keep production agents running long enough to complete multi-day work. AgentCore runtime instances let you run agents on dedicated Amazon EC2 compute, including GPU-accelerated, memory-optimized, and compute-optimized instances, with sessions lasting up to 14 days. AgentCore also expanded to US West (N. California) and Asia Pacific (Hyderabad). You can now hand agents long-running research, coding, and monitoring work, match each workload to the compute it needs, and run it closer to the users and systems it serves.
Let agents act more independently while keeping their behavior and spending within defined boundaries. Temporal policies evaluate each action against what an agent has already done, letting you enforce sequences, prerequisites, approval gates, matching values between calls, and data freshness. Rate limiting controls requests, inference tokens, and concurrent connections by user or group. AgentCore payments lets agents access and pay for APIs, MCP resources, and paid content with infrastructure-enforced spending limits and end-to-end observability. Together, these updates let you give agents more independence while controlling what can happen, in what order, at what rate, and within what budget.
Give agents timely context without weakening data boundaries. Web Search in AgentCore lets agents include or exclude domains and filter results by publication date, helping them retrieve current information from approved sources. AgentCore memory can extract long-term memories from activity logs, behavioral events, system events, and other structured JSON data, not only from conversations. Fine-grained access control can then isolate those memories by user or tenant. Now you can ground agents in trusted sources, build memory from business events, and keep that context isolated without adding custom authorization logic to every application.
Find and reuse approved agents and tools instead of rebuilding what already exists. AWS Agent Registry gives you a searchable, governed catalog for agents, MCP servers, skills, and custom resources. Organization-wide detection can identify agents on AgentCore Runtime and MCP servers on AgentCore Gateway across connected AWS accounts, while approved resources can also surface in Amazon Quick. You can find trusted capabilities by intent or name, reuse them across workflows, and reduce agent sprawl and duplicative development.
Build with more choice in regulated environments
Bring advanced models, agent capabilities, and cross-modal retrieval to regulated workloads. Claude Opus 5 is now available in AWS GovCloud (US) Regions with zero data retention enabled by default. OpenAI GPT-5.6 Terra and Luna are also available in AWS GovCloud (US) Regions, with million-token context windows and prompt caching. Amazon Nova Multimodal Embeddings brings retrieval across text, documents, images, video, and audio to AWS GovCloud (US-West), while AgentCore memory, policy, and the managed harness add managed context, controls, and orchestration. This helps you build coding agents, document-analysis systems, multimodal retrieval applications, and governed agent workflows in AWS GovCloud using more of the same capabilities available in commercial Regions.
Extend AI agents into the physical world
Move from robot demonstrations to training and physical deployment through one connected workflow. Strands Robots connects Strands Agents, LeRobot, and Hugging Face Storage Buckets, letting you record demonstrations, stream datasets in the LeRobot format, train policies, and deploy them to simulated or physical hardware without converting the underlying data. This allows you to iterate on robotics policies using one data format and one agent workflow from simulation through physical deployment.
Coordinate multiple robots and devices without rebuilding the agent for each environment.Strands Robots supports mesh-based discovery and coordination across simulated and physical devices. Zenoh connects robots on the same local network, while AWS IoT Core supports geographically distributed fleets. Strands Robots and AWS are also participating in the limited research preview of the Model Hardware Standard. This makes it easier for you to prototype multi-robot workflows in simulation, move them across local and cloud-connected fleets with fewer code changes, and explore standardized controls for physical equipment.
Get started
Explore Amazon Bedrock, deploy agents with the AgentCore CLI, or build your first agent with the Strands Harness SDK.
Interested in learning how Amazon Bedrock can support your team?Connect with us to start the conversation.****