11-19 更新 实时
首页/ 算力/ 正文

用下一代 AgentCore 运行时安全高效运行智能体

算力 来源:AWS Machine Learning 2026-11-13 01:59 0 阅读 机器翻译
用下一代 AgentCore 运行时安全高效运行智能体

AI 智能体已经不再局限于一次性的提示词。它们现在可以连续数小时运行多步骤的工作负载。前沿智能体能够独立完成的任务长度大约每四个月翻一番(METR),同时有 40% 的大型企业正在规模化部署 AI 智能体,高于一年前的 27%(McKinsey)。随着智能体运行时间更长、自主性更高,关注点正从智能体能做什么,转向它们如何以安全、可靠且经济的方式大规模运行。智能体所做的已超出传统软件的范畴:它们执行代码、操作浏览器、访问敏感数据,并代表用户采取行动。这些能力带来了新的安全和基础设施需求,使得强隔离成为生产部署的必要条件。

每个智能体会话都需要一个安全、隔离的环境。但要在规模化条件下实现这种隔离,会带来一个根本性挑战:在性能、成本与效率之间取得平衡。预置环境即使在空闲时也会占用资源,从而推高成本。按需环境减少了容量浪费,却引入了冷启动延迟,拖慢智能体的执行。如果不做取舍,这两种方式都无法高效扩展。随着企业从部署少量智能体转向同时运行数千个会话,构建和管理自定义沙箱正变得越来越复杂、昂贵且难以预测。

挑战显而易见。只有 15% 的组织已经规模化部署了经过编排的多智能体系统,而 70% 的组织表示他们目前还无法信任和治理自己的智能体(Deloitte)。扩展智能体不应意味着在安全、速度和成本之间做取舍。下一代智能体基础设施必须提供强隔离、快速启动、弹性扩展和按用量计费的经济性。因为随着 AI 智能体进入生产环境,安全执行不应成为瓶颈,而应成为让规模化成为可能的基础。通往生产环境的最快路径,也应该是最安全的路径。

我们构建 Amazon Bedrock AgentCore,是为了帮助开发者安全地大规模构建、连接和优化智能体。AgentCore runtime 是 Amazon Bedrock AgentCore 的一项能力,属于托管计算层。它为开发者提供完全托管的环境,无需构建或维护基础设施即可部署和运行智能体。自发布以来,我们的客户一直在将智能体推向日益复杂的工作负载。随着这些工作负载增长,对执行环境的要求也随之提高:更多算力、更灵活的计费方式,以及不仅覆盖计算、还延伸到数据的信任边界。只有当底层基础设施同时为规模化和速度而构建时,智能体才能实现规模化生产落地。

今天,我们宣布 AgentCore runtime 的更新,以满足这些需求。核心保证依然不变:每个会话都在其独立的硬件隔离 Firecracker 微型虚拟机(MicroVM)中运行,拥有自己的身份、策略和审计追踪。这些更新在这一保证范围内拓展了可能性。现在,会话启动速度最高提升 15 倍,可扩展至 4 倍的算力,运行效率最高提升 68%,且隔离边界完全保持完整。

以下是新增内容。

构建日益先进智能体的团队提出了更高的灵活性要求:能够在隔离会话内扩展内存和算力,以匹配其工作负载的需求。从大型代码库编译到浏览器驱动的工作流,再到本地模型推理,执行环境都需要跟上节奏。

与上一版本相比,会话现在可扩展 4 倍,最高支持 8 个 vCPU 和 32 GB,并支持最大 10 GB 的容器镜像。会话启动时间不足 2 秒,且不受镜像大小和并发量的影响。你不再需要预热池和预置集群来管理冷启动。

“迁移到下一代 AgentCore 运行时后,我们的智能体会话冷启动从近一分钟缩短到不足五秒,即使在流量高峰期也是如此。它还消除了超过 30 秒的启动延迟,而同期我们的每周智能体调用量翻了一倍多。”

— Siwei Shen,Coinbase 工程副总裁

“Mission 正在使用 Amazon Bedrock AgentCore 运行时来驱动一个自然语言查询智能体,帮助用户无需穿梭于复杂的数据系统,即可获取房产与销售洞察。在我们对两个运行时版本的初步对比中,同样的工作负载使推理请求的冷启动快了 35%,健康检查的启动快了 77%,同时将实测的单次调用运行时成本降低了 17%。在我们把企业级智能体投入生产的过程中,新的 AgentCore 运行时为我们带来了更快的智能体体验与更高效的资源利用这一极具吸引力的组合。”

— Ryan Ries 博士,Mission Cloud 首席 AI 与数据科学家

“我们的做法是,尽可能用自动化和 AI 智能体来运行 Zuora 的云,而这只有在运行时高效且可靠的前提下才成立。新一代 AgentCore 运行时在众多用例中将我们每次调用的内存占用削减了约 60%,同时以稳定的延迟处理更多流量。这让我们在平台上扩展智能体式工作负载时,拥有了更强的成本结构。”

—— Zuora 云与安全副总裁 Brad Ruppert

“Rescale 平台通过智能体式数字工程,帮助研发与工程团队加速产品开发。我们面向工程场景的专用智能体会校验仿真输入、排查作业故障、选择硬件并生成报告,而工程师始终掌握控制权。Amazon Bedrock AgentCore 运行时让我们无需构建或运维底层的智能体基础设施,就能将这些智能体部署到生产环境并实现扩展,因此我们能更快地向客户交付新的智能体。”

—— Rescale 工程副总裁 Mark Whitney

随着各团队把智能体推向生产环境,他们要求计费方式能够反映智能体的实际工作方式。智能体的大量时间都花在等待模型响应、工具调用或人工决策上。智能体应当为其完成的工作付费,而不是为其等待所耗费的时间付费。

该运行时现在能在智能体释放内存后的几分钟内回收内存,空闲的智能体可缩容至零。对于交互之间存在自然停顿的多轮智能体,这意味着每次会话的内存消耗最多可减少 68%。对于常驻在线的智能体,承诺基准定价为每次会话设定内存基准值,按最低用量计费,并可按需突发至该基准值的四倍。这为稳定负载提供了可预测的费率,与 AWS Lambda MicroVM 的费率一致。该运行时更高效地利用算力,因此同样的工作量你支付的费用更少。这让往往成为生产部署门槛的财务沟通更容易达成。

“新一代 Amazon Bedrock AgentCore 运行时通过更快启动智能体、更高效地使用内存,帮助我们交付响应更快的 AI 体验。让消费与实际用量相匹配,使我们的客户有信心扩展智能体式解决方案,同时让成本与实际需求挂钩。”

——Deloitte Consulting LLP 董事总经理、AI 与数据负责人 Ashwin Patil

“AgentCore 运行时为我们提供了一个托管基础,让我们能围绕客户的数据、工具和业务流程构建智能体产品和平台。新一代 AgentCore 运行时通过恢复已初始化的环境和回收未使用的内存,使这一基础响应更快、资源效率更高。我们正开始基于最新版本进行构建,因为成本和延迟的降低对我们客户的业务运营方式产生了显著影响。”

——Caylent 首席技术官 Randall Hunt

许多企业的智能体安全模型都存在一个长期缺口。智能体运行在隔离会话内部,但它所需的数据却位于该边界之外。长时间运行的智能体会在使用过程中不断积累上下文:文件、凭证、工具输出等等。团队最终只能自行制定数据进出的规则,每新增一个智能体,都会带来更多复杂性和风险。

仅止步于计算层面的隔离只是半个边界。下一代 AgentCore 运行时补齐了这一缺口。每个会话仍然运行在各自硬件隔离的环境中,拥有独立的身份、访问策略和审计轨迹。智能体现在可以动态挂载文件系统并保存其工作成果,例如支持 FUSE 的 Amazon Simple Storage Service(Amazon S3)存储桶。在整个任务期间,你的数据始终处于信任边界之内,由你一次性定义的一套控制措施统一管理。

“在 Okta,AgentCore 运行时与 Lambda MicroVMs 的组合提供了我们的智能体所需的隔离执行环境和上下文记忆,使其能够在临时沙箱中检查代码、运行测试套件并提交拉取请求。配合我们由身份治理的策略控制,AgentCore Runtime 和 Lambda MicroVMs 展示了我们如何安全地扩展自主工程智能体,而不会带来工作区泄露或权限提升的风险。”

—— Monica Bajaj,Okta 工程高级副总裁

原文较长,译文为节选;译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(Run agents securely and efficiently with the next generation AgentCore runtime)

AI agents have moved past one-off prompts. They now run multi-step workloads for hours at a time. The length of task that frontier agents can complete on their own doubles about every four months (METR), and 40 percent of large enterprises are scaling AI agents, up from 27 percent a year ago (McKinsey). As agents run longer and with more autonomy, the focus shifts from what agents can do to how securely, reliably, and economically they can operate at scale. Agents do more than traditional software: they execute code, navigate browsers, access sensitive data, and act on behalf of users. These capabilities introduce new security and infrastructure demands, making strong isolation essential for production deployments.

Every agent session needs a secure, isolated environment. But delivering that isolation at scale introduces a fundamental challenge: balancing performance, cost, and efficiency. Pre-provisioned environments consume resources even when idle, driving up costs. On-demand environments reduce wasted capacity but introduce cold-start latency, delaying agent execution. Neither approach scales efficiently without trade-offs. As organizations move from deploying a handful of agents to running thousands of concurrent sessions, building and managing custom sandboxes becomes increasingly complex, expensive, and unpredictable.

The challenge is evident. Only 15 percent of organizations have scaled orchestrated, multi-agent systems, and 70 percent say they can’t yet trust and govern their agents (Deloitte). Scaling agents shouldn’t mean choosing between security, speed, and cost. The next generation of agent infrastructure must deliver strong isolation with fast startup times, elastic scaling, and usage-based economics. Because as AI agents move into production, secure execution shouldn’t be the bottleneck. It should be the foundation that makes scale possible. The fastest path to production should also be the safest one.

The next generation AgentCore runtime

We built Amazon Bedrock AgentCore to help developers build, connect, and optimize agents securely at scale. AgentCore runtime, a capability of Amazon Bedrock AgentCore, is the managed compute layer. It gives developers a fully managed environment to deploy and run agents without building or maintaining infrastructure. Since launching, our customers have been pushing their agents into increasingly complex workloads. As those workloads grow, so do the demands on the execution environment: more compute, more flexible billing, and trust boundaries that extend to data, not only compute. Agents reach production at scale if the infrastructure underneath was built for both scale and speed.

Today we’re announcing updates to AgentCore runtime that meet those demands. The core guarantee stands: every session runs in its own hardware-isolated Firecracker micro virtual machine (MicroVM), with its own identity, policy, and audit trail. These updates expand what’s possible within that guarantee. Sessions now start up to 15x faster, scale to 4x more compute, and run up to 68% more efficiently, with the isolation boundary fully intact.

Here’s what’s new.

Run bigger workloads without waiting longer to start

Teams building increasingly advanced agents asked for more flexibility: the ability to scale memory and compute within the isolated session to match the demands of their workloads. From large codebase compilations to browser-driven workflows to local model inference, the execution environment needs to keep pace.

Sessions now scale 4x compared to the previous version, up to 8 vCPUs and 32 GB, with support for container images up to 10 GB. Sessions start in under 2 seconds, independent of image size and concurrency. You no longer need warm pools and pre-provisioned clusters to manage cold starts.

“Migrating to the next generation of AgentCore runtime reduced our agent session cold starts from nearly a minute to under five seconds, even during peak traffic. It also eliminated startup delays longer than 30 seconds, all while our weekly agent volume more than doubled.”

— Siwei Shen, VP of Engineering, Coinbase

“Mission is using Amazon Bedrock AgentCore runtime to power a natural-language query agent that helps users retrieve property and sales insights without navigating complex data systems. In our initial comparison of both runtime versions, the same workload delivered 35% faster cold starts for inference requests and 77% faster starts for health checks, while reducing measured runtime cost per call by 17%. The new AgentCore runtime gives us a compelling combination of faster agent experiences and more efficient resource utilization as we bring enterprise agents into production.”

— Dr. Ryan Ries, Chief AI & Data Scientist, Mission Cloud

“Our approach is to run Zuora’s cloud with automation and AI agents wherever we can, and that only works if the runtime is efficient and reliable. The next generation AgentCore runtime cut our memory footprint per invocation by around 60% in many use cases while handling more traffic at steady latency. That gives us a stronger cost profile as we expand agentic workloads across our platform.”

— Brad Ruppert, VP of Cloud and Security, Zuora

“The Rescale platform helps R&D and engineering teams accelerate product development with agentic digital engineering. Our engineering-specific agents validate simulation inputs, troubleshoot jobs, select hardware, and generate reports, while engineers stay in control. Amazon Bedrock AgentCore runtime allows us to deploy and scale those agents in production without building or operating the underlying agent infrastructure, so we can deliver new agents to customers faster.”

— Mark Whitney, VP Engineering, Rescale

Stop paying for idle agents

As teams move agents closer to production, they asked for billing that reflects how agents actually work. Agents spend much of their time waiting on a model response, a tool call, or a human decision. Agents should be billed for the work they do, not the time they spend waiting.

The runtime now reclaims memory within minutes of the agent releasing it, and an idle agent scales to zero. For multi-turn agents with natural pauses between interactions, this means up to 68 percent less memory consumed per session. For always-on agents, committed baseline pricing sets a memory baseline per session, billed as a minimum, with room to burst up to four times that baseline on demand. This gives steady workloads a predictable rate at AWS Lambda MicroVM rates. The runtime uses compute more efficiently, so you pay less for the same work. That makes the finance conversation that often gates production rollouts easier to close.

“The next generation Amazon Bedrock AgentCore runtime helps us deliver more responsive AI experiences by starting agents faster and using memory more efficiently. Aligning consumption with actual usage gives our clients great confidence to scale agentic solutions while keeping costs connected to real demand.”

— Ashwin Patil, Principal, AI and Data Leader, Deloitte Consulting LLP

“AgentCore runtime gives us a managed foundation for building agentic products and platforms around our customers’ data, tools, and business processes. The next generation Agentcore runtime makes that foundation more responsive and resource-efficient by restoring initialized environments and reclaiming unused memory. We’re beginning to build on the latest version because the cost and latency reductions make a significant difference in how our customers’ businesses operate.”

— Randall Hunt, CTO, Caylent

Isolation now extends to your data

Many enterprises face a persistent gap in their agent security model. The agent runs inside an isolated session, but the data it requires resides outside that boundary. Long-running agents pick up context as they go: files, credentials, tool outputs, and more. Teams end up building their own rules for how data moves in and out, adding complexity and risk with every new agent.

Isolation that stops at compute is only half a boundary. The next generation AgentCore runtime closes that gap. Each session still runs in its own hardware-isolated environment with its own identity, access policy, and audit trail. Agents can now dynamically mount and save their work to a filesystem, such as an Amazon Simple Storage Service (Amazon S3) bucket with FUSE support. Your data stays inside the trust boundary for the entire job, governed by a single set of controls you define once.

“At Okta, the combination of AgentCore runtime and Lambda MicroVMs provide the isolated execution environment and contextual memory our agents need to inspect code, run test suites, and open pull requests within ephemeral sandboxes. Paired with our identity-governed policy controls, AgentCore Runtime and Lambda MicroVMs demonstrate how we can safely scale autonomous engineering agents without risking workspace leakage or privilege escalation.”

— Monica Bajaj, SVP Engineering, Okta

“From helping car shoppers value a trade-in to helping dealers book service appointments and structure deals, Amazon Bedrock AgentCore powers dozens of agentic solutions across Cox Automotive’s brands. Having a single, governed agent platform lets our teams across our business move from pilot to production with confidence.”

— Chris Dillon, VP, Product and Engineering, Cox Automotive

Our commitment

The agents that reach production are the ones you trust to act on your behalf, with real data, real stakes, and real accountability. Trust isn’t a feature you add later. It’s the runtime you build on. This release is a step toward that future: bigger sessions, efficient resource use, and storage inside the trust boundary, without loosening the isolation that makes the work trustworthy. By re:Invent, per-session scoping will bind each session to its own storage folder through session-labeled credentials, and we will keep building from there.

Get started

The next generation AgentCore runtime is available today in AWS Regions across North America, South America, Europe, and Asia Pacific. Visit Amazon Bedrock AgentCore to build on it or open the Amazon Bedrock console to start exploring.


About the author

Madhu Parthasarathy

Madhu Parthasarathy

Madhu is the GM of Amazon Bedrock AgentCore, where he leads the team building the platform that companies use to build, connect, and optimize production AI agents. He brings more than 20 years of experience building large-scale distributed infrastructure, including over 16 years at Amazon, where he has led major initiatives across Amazon Retail, Amazon Elastic Block Store (Amazon EBS), and now AgentCore. Before returning to Amazon, Madhu held senior leadership roles at LinkedIn, where he led the enterprise platform powering all of LinkedIn’s enterprise lines of business, and at a neo-cloud startup, where he led AI infrastructure and set the vision for security and developer experience. He is based in Santa Clara, California.

正文保留英文原文,标题与摘要由机器翻译。

本文由采集器自动同步

原文发布于 AWS Machine Learning(Run agents securely and efficiently with the next generation AgentCore runtime),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗