过去，人们与大模型的交互是一条相对简单的链路：输入 Prompt，模型生成 Response。进入 Agent 时代，这条链路变成了“规划—推理—行动—反馈”的循环。一次用户请求，可能进一步调用网页、API、数据、工具、代码和其他系统，展开为成千上万个机器动作。
以“分析销售数据并生成报告”为例，Agent 需要读取文件、准备依赖、编写并运行代码、检查结果，再根据反馈修改重试。生成文字只是其中一步，任务要靠一连串实际操作才能完成。CPU 也因此不再只是为 GPU 做配套，而要参与整个工作流的编排和 Action 执行。
问题随之而来：当 AI 从“生成答案”走向“执行任务”，过去围绕模型推理设计的基础设施，还够用吗？腾讯云与英特尔围绕智能体基础设施的协同，正是从这条更长的执行链路出发，回应计算、内存与执行环境上的三重挑战。
第一重挑战：算力有起伏，现场不能丢
Agent 并不是始终满负荷计算。它更像“想一下，干一下，再想一下”：等待模型或工具返回时，CPU 需求回落；处理数据、执行代码、运行测试时，计算需求又迅速上升。这是 Agent 负载的常态——资源需求随任务推进起起伏伏，要求基础设施能够跟随行动节奏灵活调配算力，而不是始终按照峰值预留。
计算可以暂歇，工作现场却不能随之消失。已经载入的数据、装好的依赖、执行到一半的程序，以及多轮交互中累积的状态，都需要跨越多个 Action 持续保留。第一重挑战由此出现：：应对突发性计算负载（bursty workloads），同时保障任务状态的持久化（state persistence），让 Agent 在下一次行动时能够接着干。
按峰值长期预留资源，等待阶段容易闲置；每次空闲都拆掉环境，下一步又要重新搭建现场。Agent 需要的弹性基础设施，不能只做到“忙时加核、闲时收核”，还要能够保存现场，并在新的行动到来时恢复执行。弹性的对象，不只是算力，也包括支撑任务推进的状态。
腾讯云 Cube Sandbox 把这一需求拆成两种弹性。计算弹性，强调快速启动、高密度和大规模并发；状态弹性，强调暂停与恢复、快照与克隆，以及跨节点恢复。平台不再只管理一次请求需要多少资源，还要管理一段任务如何等待、接续和分支执行。
这也明确了双方协同的方向：Cube Sandbox 在运行时层面管理环境的创建、保存和接续，英特尔至强在底层支撑环境编排、任务执行与状态处理。当越来越多的 Agent 需要同时保留工作现场，软硬件配合要解决的下一个问题，就是这些现场究竟放在哪里、如何高效容纳。
第二重挑战：核心够用，内存可能先不够
增加 CPU 核心，并不必然带来同比例增长的 Agent 承载能力。在高密度执行平台上，内存既要服务正在运行的程序，也要承载需要继续保留的状态。当大量环境同时存在，制约密度的可能先是内存容量，而不是核心数量。算力能否充分利用，也取决于这些核心能否获得足够的内存。
英特尔数据中心集团软件技术总监胡勇以一组示例配置进行了测算：假设每个 Agent 占用约 2GB 内存，每个 vCPU 承载 4 个 Agent，则每个 vCPU 需要约 8GB DRAM；而在一套 256 vCPU、12 内存通道、每条 DIMM 128GB 的配置中，平台平均可提供约 6GB DRAM/vCPU，由此形成约 2GB/vCPU 的容量缺口。面向这一问题，可以从三个互补方向提升容量效率。
1. 让现有内存更高效，英特尔 IAA（存内分析加速器）可承担内存压缩与解压。在此次测试配置下，软件 LZ4 方案与 IAA Deflate 方案均达到约 2.1∶1 的压缩比；相较软件方案，IAA 将压缩延迟从 5.70 毫秒降至 0.51 毫秒，将解压延迟从 0.96 毫秒降至 0.08 毫秒，分别约为原来的 1/11 和 1/12。
2. 扩大本地 DRAM 容量，在支持的服务器上，从每通道一条内存的 1DPC 扩展到两条的 2DPC，直接增加本地容量。
3. 引入 CXL 内存扩展路径，为访问频率较低、但仍需保留的状态提供空间，减轻本地 DRAM 的压力。
压缩、扩容、分层，具体如何组合，取决于硬件配置、软件接入和真实工作集。但方向已经清楚：Agent 时代，CPU 的竞争正在从“有多少核心”，延伸到“每个核心能获得多少高效内存”。而当资源准备就绪，Agent 还缺最后一样东西：一个可以放心动手的地方。
第三重挑战：执行环境必须同时做到快、稳、密
当 Agent 开始运行代码、读写文件、调用 API、使用工具和访问系统，就不能让生成的代码直接在宿主机上随意执行。Sandbox，也就是沙箱，成了 Agent 的“工作间”。这是第三重挑战，也是三重挑战中最直接决定 Agent 能否干活的一环——这个工作间既要提供清晰的执行边界，也要及时交付，不能让任务长时间卡在准备环境这一步。
Agent 时代的沙箱，要同时满足三个要求：Fast，启动必须快；Isolated，隔离必须可靠；Dense，一台服务器必须能同时运行足够多的沙箱。腾讯云 Cube Sandbox 正是围绕这三点设计的，代表性指标包括约 48 毫秒启动、每个沙箱拥有独立 Guest Kernel，以及每实例低于 5MB 的额外开销。
没有 Sandbox，Agent 不敢干活；Sandbox 起得不够快、开得不够多，Agent 又干不了多少活。为了快，Cube Sandbox 通过内存快照恢复，避免每次创建环境都重复经历传统虚拟机的完整冷启动。内核引导、初始化和运行时加载等准备工作，不必在每次工具执行前从头再来。软件层面先省去重复准备，底层平台再高效完成必要的调度、恢复和执行。
为了多，Cube Sandbox 采用“共享只读基础+写时复制”：多个沙箱共享只读镜像和内存基础，只有各自修改过的 Dirty Pages，也就是脏页，才需要单独保存。好比许多 Agent 共用一套备好的工作台，各自另存自己写下的内容，而不必把整间工作室复制一遍。
这也是对内存挑战的软件侧回应：减少环境之间的重复占用，为真实增长的任务数据留出更多空间。但启动机制高效、内存复制更少，还不等于高并发下就能及时交付环境。这正是双方联合测试进一步关注的问题。
联合实测：在目标时延内接住更多请求
智能体 AI 要高效运转，执行环境也要随叫随到。当工作负载开始大规模落地，平台面对的不只是创建一个沙箱，而是在极短时间内批量创建并拉起大量沙箱，高效完成执行环境的编排与隔离。能否在并发请求增加时守住目标时延，成为检验基础设施能力的关键一环。
联合测试显示，随着启用核心数从 48 核提升至 144 核，Cube Sandbox 的并发创建量从 100 增加至 250，而平均创建时延始终保持在约 200 毫秒，体现出良好的并发扩展能力。进一步测试中，在平均创建延迟 200 毫秒的条件下，搭载至强®6+处理器的 Cube Sandbox 可支持 350 个沙箱并发创建请求，并发密度达到其他主流测试平台的 1.46 倍。
对于 Agent 基础设施而言，真正重要的并不只是单个沙箱启动有多快，而是当大量 Agent 同时“开工”时，平台能否在目标时延内持续交付足够多的执行环境。这也让 CPU 的价值衡量从单任务性能，进一步走向面向 Agent 执行的系统级并发效率。
在这一过程中，Cube Sandbox 通过快照恢复、共享基础镜像和写时复制等机制，减少重复初始化与内存复制；英特尔® 至强® 6+则以高密度计算和平台扩展能力支撑大规模沙箱的并发创建与执行。二者的软硬件协同，最终指向一个更贴近 Agent 实际应用的指标：
在既定时延目标下，一台服务器究竟能让多少 Agent 同时“开工”？
实测结果也让处理器性能与 Agent 规模化部署真正建立起联系。当大量沙箱创建请求集中到来时，平台不仅需要“有资源”，更需要及时将资源转化为可用的执行环境。这也意味着，未来 Agent 基础设施比拼的将不只是单点性能，更是面向规模化执行的持续交付能力。
状态接续：让工作间成为 Agent 运行时
环境就绪之后，Agent 还要在受控边界内持续工作。Cube Sandbox 采用 MicroVM 架构，为每个沙箱提供独立 Guest Kernel，底层依托至强处理器与 KVM 硬件虚拟化能力。独立内核与虚拟化机制共同构成执行边界，避免 Agent 生成的代码直接运行在宿主机环境中。
网络侧，Cube Sandbox 采用 eBPF 隔离，默认拒绝、按白名单放行；凭据由七层代理在出站请求中注入，不直接暴露给沙箱内代码。独立内核、网络策略和凭据管理分别作用于不同环节，使环境隔离与外部访问控制相互配合，为 Agent 执行工具和访问系统提供防护。
但沙箱的价值并不止于隔离。Cube Sandbox 还提出像 Git 一样管理环境：运行中的沙箱可以保存快照，再克隆出多个分支，分别继续执行、失败后丢弃，或者暂停后再恢复；借助跨节点恢复，状态也不必始终绑定原来的计算节点。第一重挑战提出的“状态弹性”，在这里变成了具体能力。
对于需要反复试错、跨轮交互的 Agent，这意味着环境可以随着任务进展被保存和接续。暂停不必等于清空现场，尝试另一条执行路径也不必每次从零开始。沙箱由此从安全隔离环境，进一步成为 Agent Runtime 的重要执行基础，支撑任务在多个行动之间持续推进。
从 Cube Sandbox 管理环境生命周期，到至强提供计算、虚拟化与状态处理支持，双方合作将三重挑战连成了一条线。计算需要按需调配，内存需要高效承载，执行环境则要及时就绪、受控运行并保留现场。只有这些环节衔接起来，Agent 才能从一次模型调用走向一段完整的工作过程。
从 Scale Inference 走向 Scale Execution
模型中心的 AI，链路是“Prompt—Model—Response”，核心问题是 Scale Inference，关注模型吞吐与推理效率；Agentic AI 则进一步走向 Scale Execution，关注整个任务能否安全、高效、大规模地完成。基础设施的衡量尺度，也从每秒生成多少 Token，扩展到能够及时交付多少执行环境、支撑多少任务持续推进。
腾讯云 CubeSandbox 与英特尔® 至强® 6+ 的合作，为这一变化提供了具体注脚。CubeSandbox 提供弹性、可信且具备状态管理能力的执行环境，至强® 6+则以高性能计算与平台扩展能力，支撑大规模沙箱的并发编排与执行。双方的联合验证，让软硬件协同从架构设计落实到实际效果，体现为目标时延内更强的沙箱并发创建能力，充分体现了至强® 6+在 Agentic AI 领域的优势。
当 AI 从“会想”走向“会干”，CPU 的价值也开始从支撑模型，走向支撑海量 Agent 真正行动起来。至强® 6+与 Cube Sandbox 的协同，让执行环境更快就绪，让更多 Agent 同时“开工”，也进一步展现了 CPU 作为 Agentic AI 执行底座的价值。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
In the past, people interacted with LLM as a relatively simple link: input Prompt, model generation Response. In the age of Agent, the link became a cycle of “planning-debit-action-revenue”. A single user request may further call web pages, API, data, tools, codes and other systems for thousands of machine actions.
Using the example “Analysis of sales data and generation of reports”, Agent needs to read documents, prepare to rely on them, prepare and run codes, check results, and modify retry based on feedback. Generating text is only one step in the process, and the task is done by a series of practical operations. CPU is therefore no longer just a GPU companion, but is involved in the entire workflow organization and action implementation.
The question follows: When AI moves from “generation of answers” to “tasking,” is the infrastructure that used to be designed around model reasoning enough? The synergy between the Tencent cloud and Intel around the AI agent infrastructure proceeds from this longer implementation chain and responds to the triple environmental challenge of computing, memory and implementation.
The first challenge is to have an ups and downs and downs.
Agent is not always full. It is more like "Think about it, dry about it, think about it": waiting for a model or tool to return, the CPU needs fall; and when processing data, implementing codes, running tests, the computing needs rise rapidly. This is the normal pattern of the Agent load — resource needs that fluctuate as the mandate moves, requiring that the infrastructure be able to deploy the calculus in a flexible manner following the pace of action, rather than always set at peaks.
The calculation can be suspended and the work site cannot disappear. The data already contained, the installed dependency, the procedures to be carried out in half, and the cumulative status of multiple rounds of interaction all need to be kept over multiple Action on a continuous basis. The first challenge follows: responding to sudden-onset calculations (bursty workloads) while safeguarding the sustainability of the mission to enable Agent to continue with the next operation.
Resources are set aside for a long period of time at peaks, and the waiting phase is easily idle; each free time the environment is dismantled and next step re-establishes the site. Agent requires a flexible infrastructure that cannot be “quicker, freer” only, but also capable of keeping the site and resuming implementation when new operations arrive. Resilient objects are not just calculators, but also the state of support for mission advancement.
Tencent Cloud Cube Sandbox has broken this demand into two types of elasticity. Computative elasticity, with emphasis on rapid start-up, high-density and large-scale co-opt; state elasticity, with emphasis on suspension and restoration, snapshot and cloning, and recovery across nodes. The platform no longer manages the amount of resources required for a single request, but also how a mandate is to wait, continue and be carried out.
It also provides a clear direction for synergies: Cube Sandbox manages the creation, preservation and continuation of the environment at the operational level, and Intel to Power supports environmental organization, task execution and state of operation at the bottom. As more and more Agents need to keep their work sites at the same time, the next question to be addressed with hardware and software is where they are located and how they can be accommodated efficiently.
Second challenge: Core is adequate and memory may not be sufficient first
The addition of the CPU core does not necessarily result in the growth of the Agent carrying capacity on a proportional basis. On a high-density implementation platform, memory is required both to service the running process and to carry the state that needs to be retained. When a large number of environments coexist, limiting density may begin with memory capacity rather than core quantity. Optimization also depends on the availability of sufficient memory for these cores.
The Director General of Software Technology of the Intel Data Centre Group, Hu Yong, measured with a set of illustrative configurations: assuming each Agent occupied approximately 2GB memory and each vCPU carried 4 Agent, each vCPU required approximately 8GB DRAM; and in a set of 256 vCPU, 12 memory channels, each DIM 128GB configuration, the platform provided an average of about 6GB DRAM/vCPU, resulting in a capacity gap of about 2GB/vCPU. In this context, capacity efficiency can be enhanced in three complementary directions.
1. To make the existing memory more efficient, the Intel IAA can carry the memory compression and decompression. Under this test configuration, both the LZ4 program and the IAA Deflate program have reached a compression ratio of about 2.1:1; compared to the software program, IAA reduced the compression delay from 5.70 ms to 0.51 ms and the decompression delay from 0.96 ms to 0.08 ms, about 1/11 and 1/12, respectively.
2. Expand the local DRAM capacity from one memory of 1DPC to two 2DPCs on a supported server, directly increasing local capacity.
3. Introduction of CXL memory extension paths to provide space for access to less frequent states that still need to be retained, and to reduce local DRAM pressure.
Compression, amplification, layering and how they are combined depend on hardware configuration, software access and real work sets. But the direction is clear: in the Age of Agent, the CPU competition is extending from “how many cores” to “how many efficient memory per core”. And when the resources are ready, Agent needs the last thing: a place to do it.
Third challenge: the implementation environment must be fast, secure and secure at the same time
When Agent starts running codes, reading and writing files, calling API, using tools and accessing systems, it is not possible to let the resulting code be implemented directly on the host. Sandbox, the sandbox, became the "working room" for Agent. This is the third challenge, and it's the most direct part of the triple challenge that determines Agent's ability to work. This workshop should provide a clear operational boundary and be delivered in a timely manner, and should not allow the task to be stuck at this stage of preparation.
The Agent sandboxes meet three requirements at the same time: Fast Start, Fast Start, Isolad, Secure, and Dense, a server must be able to run enough sandboxes at the same time. The Tencent cloud Cube Sandbox was designed around these three points, with representational indicators of about 48 milliseconds of start-up, each sandbox having a separate Guest Kernel, and additional costs of less than 5MB per example.
Without Sandbox, Agent wouldn't dare to work; Sandbox didn't get up fast enough to drive enough, and Agent couldn't do much. In order to speed up, Cube Sandbox recovers from memory snapshots, and avoids repeating the complete cold start of the traditional virtual machine every time it is created. Preparatory work such as kernel guidance, initialization and loading at the time of operation does not need to be repeated from scratch before each tool is implemented. The software level saves re-preparation and the bottom platform efficiently completes the necessary movement, recovery and implementation.
For more, Cube Sandbox uses "Shared read-only base + copy when writing": multiple sandboxes share a read-only mirror and memory base, which needs to be saved in isolation only if the Ditty Pages, or dirty pages, have been modified. It's like many Agents share a set of available workstations, each of which saves their own writings without having to copy the entire studio.
It is also a software-side response to memory challenges: to reduce duplication between environments and to allow more space for real growth of mission data. However, a start-up mechanism that is efficient and has less replicability of memory does not amount to a high level and a timely delivery environment. This is a matter of further concern for joint testing.
Joint findings: additional requests within target latency
AI agent AI has to function efficiently and the implementation environment is on call. When the load begins to land on a large scale, the platform is faced not only with the creation of a sandbox, but with the creation and lifting of large quantities of sandboxes in a very short period of time, efficiently completing the organization and isolation of the implementation environment. The ability to maintain target latency at the time of the request for additionality is a key element of the infrastructure capacity test.
Joint testing shows that with the activation of the core number from 48 to 144, the co-generation of Cube Sandbox increased from 100 to 250, while the average creation of latency remained at approximately 200 ms, reflecting good co-generation and expansion. In further testing, with an average delay of 200 ms to create, Cube Sandbox, which carries up to a powerful 6+ processor, can support 350 sandboxes and send out requests for creation, with a density of 1.46 times that of other mainstream testing platforms.
For Agent infrastructure, what is really important is not just how fast individual sandboxes are activated, but whether the platform can deliver enough of the implementation environment within target latency when a large number of Agents are at the same time “starting work”. This also allows CPU value measurement to move from single-task performance to a system level oriented towards Agent implementation and to achieve efficiency.
In this process, Cube Sandbox reduces duplicate initialization and memory reproduction through such mechanisms as snapshot restoration, sharing of basic mirrors and writing reproduction; and the co-generation and implementation of large sandboxes is supported by high-density calculations and platform extensions. The software and hardware synergies of the two eventually point to an indicator closer to Agent's practical application:
How much does a server get Agent to "work" at the same time under the established latency target?
The results also make a real link between processor performance and Agent's scale deployment. When a large number of sandbox creation requests are concentrated, the platform needs not only “resources” but also the timely translation of resources into an operational environment. This also means that, in the future, the Agent infrastructure will be more than just a single-point performance, but more of a continuous delivery capacity for scale-oriented implementation.
Organisation Time
Once the environment is ready, Agent will continue to work within the controlled boundaries. Cube Sandbox uses the MicroVM structure to provide an independent Guest Kernel for each sandbox, with a bottom-to-power virtualization capability for the processor and KVM hardware. Independent kernels, together with virtualization mechanisms, constitute the enforcement boundary and avoid the direct operation of Agent-generated codes in the host environment.
On the side of the network, Cube Sandbox uses eBPF isolation, by default refuses and is released by white list; by evidence, the seven-story agent is injected in the exit request without direct exposure to the sandbox code. Independent kernels, web-based strategies and evidence-based management play different roles, aligning environmental isolation with external access controls and providing protection for Agent implementation tools and access systems.
But the value of sandboxes goes beyond isolation. Cube Sandbox also proposes to manage the environment like Git: operating sandboxes can save snapshots, then clone multiple branches, continue to run, fail and then be abandoned, or put back on hold; and, by means of cross-node restoration, state does not always have to bind the original calculation node. The “state elasticity” posed by the first challenge has been translated here into concrete capabilities.
For Agent, which needs to be tested repeatedly and interactively, this means that the environment can be preserved and continued as the mission progresses. The suspension does not have to amount to vacating the site, nor does it have to try another implementation path from scratch. As a result, sandboxes have been placed in a safe and secure environment, which has further served as an important operational basis for Agent Runtime, supporting the continuous advancement of the mandate across multiple operations.
From the Cube Sandbox management environment life cycle to the strong provision of computing, virtualization and state-of-the-art support, the three-fold challenge is linked together. The calculation needs to be aligned as needed, the memory needs to be efficiently carried, and the implementation environment needs to be ready, controlled and on-site. Only when these links are linked can Agent move from a model to a complete process.
Scale Access
AI, the model centre's link is “Prompt-Model-Response”, the core issue is Scale Reference, which focuses on model ingestion and reasoning efficiency, while Agentic AI moves further towards Scale Exchange, focusing on whether the entire mission can be accomplished safely, efficiently and on a large scale. The measure of infrastructure also extends from how much Token is generated per second to how much the environment can be delivered in a timely manner and how much the mission sustains.
Tencent Cloud CubeSandbox ' s collaboration with Intel to Power 6+ provides a specific note to this change. CubeSandbox provides a flexible, credible and well-manufactured implementation environment, and up to 6+ provides high-performance computing and platform expansion capabilities to support the co-programming and implementation of large sandboxes. The joint validation of the two parties, which allowed for a synergy of software and hardware from the design of the architecture to its practical effects, reflected a stronger co-production capacity in the sandbox within the target latency, fully reflected the advantages of up to 6+A in the field of Agentic AI.
And when AI went from "thinking" to "dry," the value of CPU started to move from a support model to one that supported the big Agent. The synergizing of up to 6+ with Cube Sandbox to make the implementation environment faster and more Agent at the same time “starts work” further demonstrates the value of the CPU as an executive base for Agenic AI.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.