国际新闻 · 中文译文与原文对照

星期五 最新 1,000 篇 · 滚动浏览

科技

谷歌向CNCF捐赠Agent Substrate和gVisor,以助力大规模安全部署AI智能体。

来源:InfoQ 技术发布 上传时间 (多伦多时间)1 家来源 · 1 篇报道

中文译文

谷歌宣布将其开发的Agent Substrate和gVisor捐赠给云原生计算基金会(CNCF),将其人工智能代理基础设施中的两个重要组件置于中立的开源管理框架之下。这一举措旨在解决大规模运行自主代理时面临的两个新兴挑战:安全地执行不可信的、由人工智能生成的代码,以及应对那些频繁暂停和恢复运行的代理所带来的巨大计算开销。

该公告还提到了Google的开源强化学习编排项目OpenRL。这些项目共同覆盖了代理基础设施的不同层面,从隔离的代码执行到代理生命周期的管理,再到大规模训练工作负载的协调。其目标是为代理应用建立开放的基础框架,就如同Kubernetes推动了容器编排的标准化一样。谷歌认为,传统的Kubernetes工作负载与人工智能代理在执行模式上存在根本差异。传统的应用程序通常会启动后持续处理请求并运行。而人工智能代理则可能在执行工具或代码几秒钟后,便需要等待模型推理结果、外部API响应或人工审批。谷歌估计,这些代理有90-95%的时间处于空闲状态,但仍能保持重要状态,包括内存、文件、存储库以及正在进行的会话。

将每个代理都放在独立的Pod中会浪费资源,而将规模缩减到零则会在需要重新构建环境时造成延迟。另一个极端是,当成千上万个短生命周期的执行环境被反复创建和销毁时,大规模的代理测试可能会让Kubernetes控制平面不堪重负。

为了解决这些限制,谷歌开发了Agent Substrate,通过将代理状态(即执行单元)与用于执行它的计算资源(即工作节点)分开来应对这些问题。演员节点代表各个独立的代理会话,并保留其状态,包括文件系统内容、内存以及执行上下文。工作节点则提供从共享的Kubernetes资源池中获取的、可重复使用的、已预热的执行环境。

当某个代理需要恢复运行时,Substrate会将其分配给一个可用的工作节点,并恢复其先前的状态。当该代理处于空闲状态时,其状态可以被保存,相应的计算资源也会被释放。

据谷歌称,这种方法能够让代理在500毫秒以内恢复运行,而无需为每次执行都创建新的Kubernetes Pod。此外,通过独立于各个Pod对象来管理代理的生命周期,这种方式还能减轻Kubernetes控制平面的负担。

第二个主要挑战是安全性。AI代理通常会执行生成的shell命令、安装软件包以及与外部工具进行交互。如果将这些代码视为可信的,就会带来重大风险,尤其是当多个代理共享同一基础设施时。Google的gVisor通过用户空间应用程序内核提供了额外的隔离边界。它不会让容器化应用程序直接与宿主Linux内核交互,而是拦截系统调用并在其自身的隔离环境中处理这些调用。

这样一来,在保持基于容器的执行方式所带来的高效性的同时,也降低了底层宿主的暴露风险。

Agent Substrate还支持用于需要硬件级虚拟化的工作负载的云管理程序微虚拟机。这为平台团队提供了灵活性,使他们能够在不改变整体代理生命周期模型的情况下选择合适的隔离边界。

这一公告反映了云原生基础设施领域的更广泛变革。Kubernetes最初是为长运行应用程序和相对稳定的工作负载设计的。而代理系统则引入了更为动态的执行模式,其中环境需要以更高的频率被创建、暂停、恢复和隔离。其他供应商也在通过Kata Containers、Firecracker以及专门的沙箱平台等技术来解决类似的挑战。谷歌的贡献在于将安全执行与具备代理感知能力的编排模型相结合,并将这些技术置于一个开放的生态系统之中。

更广泛的意义在于,人工智能代理不仅需要编排框架和工具访问权限,还需要拥有自己安全、可维护且管理高效的执行基础设施。通过将这些项目捐赠给CNCF,谷歌正在为云原生计算的下一发展阶段奠定基础。

原文

Google has announced it is donating Agent Substrate and gVisor to the Cloud Native Computing Foundation (CNCF), placing two important components of its AI agent infrastructure under vendor-neutral open-source governance. The move is intended to address two emerging challenges in running autonomous agents at scale: securely executing untrusted, AI-generated code and managing the substantial compute overhead created by agents that frequently pause and resume.

The announcement also includes OpenRL, Google's open-source reinforcement learning orchestration project. Together, the projects target different layers of agent infrastructure, from isolated code execution to managing agent lifecycles and coordinating large-scale training workloads. The goal is to establish an open foundation for agentic applications, much as Kubernetes helped standardise container orchestration.

Google argues that conventional Kubernetes workloads and AI agents have fundamentally different execution patterns.

A traditional application typically starts, serves requests, and continues running. An AI agent, however, may execute a tool or code for a few seconds before waiting for model inference, an external API, or human approval. Google estimates that agents can spend 90-95% of their time idle while still holding important state, including memory, files, repositories, and active sessions.

Keeping every agent in a dedicated Pod wastes resources, while scaling to zero introduces delays when environments must be reconstructed. At the other extreme, large-scale agent evaluations can overwhelm Kubernetes control planes when thousands of short-lived execution environments are repeatedly created and destroyed.

Google developed Agent Substrate to address these limitations by separating agent state (actors) from the compute resources (workers) used to execute it.

Actors represent individual agent sessions and retain their state, including filesystem contents, memory, and execution context. Workers provide reusable, pre-warmed execution environments drawn from shared Kubernetes pools.

When an agent needs to resume, Substrate assigns it to an available Worker and restores its previous state. When the agent becomes idle, its state can be saved and the underlying compute resource released.

According to Google, this approach lets agents resume in under 500 milliseconds without creating a new Kubernetes Pod for every execution. It also reduces pressure on the Kubernetes control plane by managing agent lifecycles independently of individual Pod objects.

The second major challenge is security. AI agents routinely execute generated shell commands, install packages, and interact with external tools. Treating this code as trusted creates significant risks, particularly when multiple agents share infrastructure.

Google's gVisor provides an additional isolation boundary through a user-space application kernel. Instead of letting containerised applications interact directly with the host Linux kernel, gVisor intercepts system calls and handles them within its own isolated environment.

This reduces the exposure of the underlying host while retaining much of the efficiency associated with container-based execution.

Agent Substrate also supports Cloud Hypervisor microVMs for workloads requiring hardware-level virtualisation. This gives platform teams flexibility to select an appropriate isolation boundary without changing the overall agent lifecycle model.

The announcement reflects a broader shift in cloud-native infrastructure. Kubernetes was designed around long-running applications and relatively stable workloads. Agentic systems introduce more dynamic execution patterns, where environments must be created, suspended, restored, and isolated at much greater frequency.

Other providers are addressing similar challenges through technologies such as Kata Containers, Firecracker, and dedicated sandboxing platforms. Google's contribution is to combine secure execution with an agent-aware orchestration model and place the resulting technologies within an open ecosystem.

The wider implication is that AI agents need more than an orchestration framework and access to tools; they need their own secure, stateful, and efficiently managed execution infrastructure. By donating these projects to the CNCF, Google is helping establish the foundations for that next layer of cloud-native computing.