智东西（公众号：zhidxcom）
作者 | 杨京丽
编辑 | 李水青
智能体的规模正在快速扩张。IDC预测，到2030年，全球活跃Agent数量将达到22.16亿个，年复合增长率达到139%，年均增长超1倍。
各类智能体也在密集涌现，热度一路走高。从编程、办公到个人助手，从消费端到企业端，智能体正快速进入各个场景，成为今年AI落地的重点方向。
与热度同步攀升的，是Token（词元）的消耗量。国家数据局披露，到今年3月，我国日均Token调用量已超过140万亿，较2024年初增长1000多倍；中国信通院的监测则显示，到今年6月，这一数字已达500万亿。如此庞大的调用量说明，智能体已经走出概念演示，开始承接大量真实任务。
这些被高频调用的智能体，大多集中在通用能力和高频场景。还有一条路线值得单独拿出来看——产业智能体。通用智能体比拼的是场景覆盖和调用频次，产业智能体则直接扎进社区治理、工厂车间、园区和门店这些业务现场，解决琐碎复杂的现实问题和日常事务。
这正是百度智能云蹚出的一条路子。在10月29日于武汉举行的百度智能云超级智能体大会上，从社区治理、园区运营、制造生产到门店管理，各类场景都涌现出百度智能体落地应用的实际案例：它可以根据居民诉求自动生成治理工单，把群众的需求变成可追踪、可闭环的处理流程；它还能将工厂老师傅的经验沉淀成可调用的技能，方便后续快速复用。
将这些实践连起来，不难看出百度智能体落地呈现出的几个特征：一是场景聚焦产业一线，治理、运营、生产、服务等环节均有实际落地案例。二是直面复杂需求，尤其是工业制造这类经验密集、流程冗长的领域，对智能体的执行能力要求更高；三是强调任务闭环，能够让需求有结果、经验可复用。
要让智能体在企业中长期、规模化地运行，需要先解决资源供给、任务执行、数据连接和安全管理这几道难题。百度智能云因此推出AI Infra与Agent Infra双层基础设施，让智能体成为企业关键生产力。
一、从个人助手到业务主力，智能体闯行业深水区
最近一年，智能体正从个人提效工具，走向企业核心业务场景，落地步伐逐步迈向深水区。会上，百度副总裁石清华称，百度搭子作为通用智能体，今年3月上线后，月活已达千万级，环比增长10倍以上，智能体正在快速向大众渗透。
在C端不断普及的同时，智能体也在企业业务场景释放产业价值。
杭州调味品企业松鲜鲜接入百度搭子后，原本需要两天完成的数据汇总工作，可缩短至3分钟。智能体展现出跨系统数据读取与业务洞察能力，开始连接企业内部系统，参与经营分析和业务判断。
企业业务场景之外，百度搭子还深度参与基层治理。在武汉硚口，政企服务企业鸿海鑫基于百度搭子的基层治理项目已覆盖100万人口、300位用户和40个真实场景。居民在网格群发布的信息，经智能体分析后可自动生成诉求工单，并推送至街道系统跟进处理。在复杂场景下，百度智能体具备非结构化数据处理能力、自动化执行能力与跨系统协作能力。
无论是企业增效还是基层治理，智能体的落地并非一蹴而就，其背后是一整套系统工程的重新构建。随着Agent在企业中的参与度不断提高，相关基础设施也需要同步升级。
百度智能云技术委员会联席主席郑然在接受智东西采访时谈道，企业调用的已不只是一个大模型接口，想要让智能体深入业务，还需要云厂商提供两类支撑：一方面是开箱即用的Agent能力，帮助企业快速接入和试错；另一方面是支撑智能体长期运行的基础设施，包括运行环境、数据连接、安全管理和持续运维等。
而这两类能力，正是百度智能云面向企业智能体落地所提供的底层支撑。
二、智能体要长效，模型之后，要先补“运行底座”
要让智能体在企业中发挥长效价值，需要为它提供一个安全稳定、价值可衡量的运行环境。
首先，智能体需要调用工具和企业内部数据，一旦操作失误就可能引发数据泄露或系统崩溃，因而需要一个安全的执行环境，“沙箱”可以让智能体在封闭环境内执行任务，避免波及其他系统。
其次，大模型还具有随机性，企业还需要知道智能体到底做了什么、花了多少成本、效果如何，这就要求基础设施具备“可观测”能力，帮助企业看清Agent成本、定位问题，推动效果持续迭代。
Agent Infra的本质，正是把一次性的模型调用，变成一个可以长期运行、可观测、可治理的任务系统。智能体部署完成后，真正困难的是让它在企业环境中稳定运行，并持续交付结果。要做到这一点，需要先了解智能体的底层技术形态。成熟的云原生应用能够按需启动，根据任务量弹性调整计算资源，通过标准化方式调用云上的计算、存储和网络资源，由平台持续完成部署、调度和运维。
郑然告诉智东西，智能体本身也是一种云原生应用，现有云原生技术仍然是其运行基础。但智能体的资源需求更具弹性，运行中也在一定的不确定性，未来还可能以海量规模运行，因此云原生基础设施需要进一步演进，以适应它的工作负载。
随着智能体数量快速增长，基础设施还要支撑更大规模的协同运行，这对底层架构提出了更高的要求。这就意味着，Agent进入生产后，企业的核心缺口逐渐从模型转向运行底座。
郑然称，企业如果从头建设Agent Infra，需要覆盖计算、存储、网络、云原生和大模型工程等多个技术领域，建设投入和持续维护成本都很高。云厂商则可以通过上下游协同和规模化运营，把这些能力沉淀为开箱即用、分层可组合的产品。企业可以根据自身所处阶段，灵活选择基础设施、Agent平台或开箱即用的智能体，快速验证想法并持续迭代。
百度智能云的Agent Infra覆盖沙箱、计算、网络、存储、可观测和安全等能力。沙箱负责快速启动任务、隔离运行环境并保存任务状态，让智能体在安全边界内执行任务；计算、网络和存储则像“水电”一样，高效支撑任务执行、数据访问和上下文处理；可观测能力记录智能体的执行轨迹，帮助企业分析成本、效果和故障原因。
通过这些能力，智能体才能长期、稳定、安全地运行。在底层能力之上，百度智能云还进一步降低了智能体调用的门槛。过去，云产品复杂的界面和繁琐的配置往往阻碍了智能体的自动化执行；为此，百度智能云专门针对Agent的调用习惯，深度优化了云产品的交互界面和使用方式，让智能体能够像“调用本地文件”一样，更高效、直接地调度云上的计算、存储和网络资源。
目前，百度智能云已在自有产品和客户项目中持续验证相关能力，为企业提供从资源供给、智能体运行到开发运营的完整支撑。
三、Token需求爆发式增长，AI Infra需要提升资源效率
Agent Infra把智能体的任务跑起来之后，背后还需要AI Infra持续提供模型训练、推理和资源供给。随着智能体数量和调用规模增长，AI Infra面对的压力也在同步上升。IDC预测，这些智能体带来的推理需求将超过100吉瓦。智能体越多、任务链路越长，算力、存储、网络和能源投入就越高，云厂商需要同时扩大资源供给，并提高每单位资源的Token产出效率。
郑然在采访中提到，智能体上下文长度不断提升，模型参数规模持续扩大，结构更加稀疏，后训练占比也在提升。这些变化对计算、存储、网络和推理效率提出了更高要求，基础设施也需要围绕模型训练与推理重新设计，提升长上下文场景下的数据访问和Token效率。
百度智能云通过百舸AI计算平台6.0，将AI计算平台从单纯提供算力，升级为能够协同基础设施、资源管理和工程能力的“智能工厂”，更高效地支撑智能体的运行。
智能体走入更多场景的背后，既需要Agent Infra承接智能体的任务执行、工具调用和运行管理，也需要AI Infra持续提升模型训练推理和资源供给效率。两者协同，才能让智能体在更长任务链路、更大规模的调用中保持稳定运行，并控制整体使用成本。
结语：Agent进入生产现场，全新云底座稳住落地需求
智能体正在从个人工具变成企业业务系统中的执行单元。模型能力只有进入生产环节，才能转化为可以持续复用的产业价值。
当前智能体在越来越多的场景中落地，不断催生新的需求，这种“新”也会带动芯片、光模块、存储等基础设施产业发展，催生新的应用场景和商业模式。
对云厂商而言，竞争重点也从提供单一算力资源，扩展到支撑智能体长期运行的完整基础设施。AI Infra与Agent Infra的协同，正在成为Agent落地的重要路径。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
YANG Kyung-Li
Editor, Lee Shui-qing
AI agent is expanding rapidly. IDC predicts that by 2030, the number of active Agents will reach 2,216 million globally, with a combined annual growth rate of 139 per cent, with an average annual increase of more than onefold.
Various types of AI agent are also on the rise and the heat is rising. From programming, office to personal help, from consumption to enterprise, AI agent is rapidly entering the scenes and becoming the focus of AI this year.
This is the consumption of Token. According to the National Data Agency, by March of this year, our average daily call to Token had exceeded 1.4 trillion, more than 1,000 times greater than at the beginning of 2024; surveillance by the China Institute of Information and Communication Technology showed that by June that figure had reached 500 trillion. Such a large call indicates that AI agent has moved out of the concept demonstration and has begun to take on a large number of real tasks.
The majority of these AI agents, which are transferred to HF, are concentrated on generic capabilities and high frequency scenes. There's another route worth looking at alone — industry AI agent. General AI agent competes with scene coverage and frequency of call, while the industry AI agent works directly into community governance, factory workshops, parks and doorshops, where complex realities and routines are addressed.
That's the way Baidu is smart. At the Baidu Smart Cloud SuperZ42Z conference in Wuhan on 29 October, from community governance, park operation, manufacturing to doorshop management, practical examples of BaiduZ42Z landing applications emerged from various scenarios: It can automatically generate a governance list based on the demands of the population, turning the needs of the population into a traceable and closed-ringed process; it can also sink the experience of factory teachers into accessible skills and facilitate subsequent quick re-entry.
The combination of these practices reveals several features of the landing of BaiduZ42Z: the first is the focus of the industry, where there are practical examples of governance, operation, production and services. The second is to confront complex needs, especially in areas such as industrial manufacturing, where experience-intensive and lengthy processes are more demanding in terms of the ability to implement AI agent; and the third is to emphasize the closed mandate, which allows demand to be productive and experience to be replicated.
For AI agent to operate in the medium- and long-term, on a scale basis, the challenges of resource availability, task execution, data connectivity and security management need to be addressed. The Baidu smart cloud thus launched the AI Infra and Age Infra two layers of infrastructure, making AI agent a key productivity for the enterprise.
I. From Personal Assistant to Business Major, AI agent breaks deep into the industry
Over the past year, AI agent is moving from a personal efficiency tool to a core business landscape for the enterprise and moving towards a deep water zone. At the meeting, Vice-President Baidu Ishi Qinghua stated that when Baidu was on the line as a General AI agent, since March of this year, the monthly life had reached a level of 10 million, the ring had increased by more than 10 times, and AI agent was rapidly penetrating to the public.
At the C end of the spectrum, AI agent releases industrial value in the business landscape.
The Baidu set, which was to have been completed in two days, could have been shortened to three minutes. AI agent has demonstrated cross-system data access and business insight, and has begun to connect to enterprise internal systems and participate in business analysis and judgement.
In addition to the business landscape, the Baidu team is deeply involved in grass-roots governance. In Wuhangun, the Baidu-based grass-roots governance project, a government-serving enterprise, has reached 1 million people, 300 users and 40 real scenes. Information on the grid group publish, analysed by AI agent, is automatically generated and sent to the street system for follow-up. In a complex scenario, Baidu AI agent has a non-structured data processing capacity, an automated implementation capacity and a cross-system collaboration capability.
Whether it is enterprise efficiency or grass-roots governance, the landing of AI agent is not a one-off event, and behind it is the rebuilding of a whole set of system works. As Agent ' s participation in business continues to increase, the infrastructure needs to be upgraded simultaneously.
The co-chair of the Baidu Smart Clouds Technology Committee solemnly spoke in interviews with intellectuals about the fact that the enterprise was no longer just a LLM interface, but that in order for AI agent to go deeper, cloud manufacturers were also required to provide two types of support: on the one hand, the Agent capability to open boxes to help enterprises to access and test; and, on the other hand, the infrastructure to support AI agent's long-term operation, including operating environments, data connections, security management and continuity.
These two types of capabilities are the bottom support provided by the Baidu smart cloud towards the landing of the enterprise AI agent.
II. AI agent is long-lived and the model is followed by a "run base"
For AI agent to be of long-lasting value to the enterprise, it needs to be provided with a secure, stable and measurable operating environment.
First, AI agent is required to call on tools and enterprise internal data, and once an operational error could trigger a data leak or a system collapse, thus requiring a safe implementation environment in which the Sandbox can enable AI agent to operate in a closed environment to avoid wave and other systems.
Secondly, LLM is random, and enterprises need to know what exactly AI agent has done, how much it has cost, and how effective it has been, which requires infrastructure to have “observable” capabilities to help businesses understand the costs and positioning of Agent and to promote continuity of effects.
The essence of Agent Infra is to turn a one-time model into a long-term, observable and manageable mission system. Once the deployment of AI agent has been completed, what is really difficult is for it to operate in a stable business environment and deliver results on an ongoing basis. In order to do this, it would be necessary to understand the technological base of AI agent. A mature cloud-based application can be initiated on a needs-based basis to adjust the computational resources to mission volume elasticity and to mobilize cloud-based computing, storage and network resources in a standardized manner, with deployment, movement control and transportation to the platform on an ongoing basis.
AI agent is also a cloud-based application, as AI agent is solemnly told, and existing cloud-based technologies continue to form the basis for its operation. However, the resource requirements of AI agent are more flexible, there is some uncertainty in its operation and it is likely that it will operate on a large volume scale in the future, so that the cloudland infrastructure needs to evolve further to meet its workload.
With the rapid increase in the number of AI agents, infrastructure will also have to support larger synergistic operations, which will place higher demands on the bottom structure. This means that after Agent entered production, the core gap of the enterprise gradually shifted from model to operating base.
For the record, if an enterprise were to build Agent Infra from scratch, it would have to cover a number of technical areas, such as computing, storage, networks, clouds and LLM works, with high construction input and ongoing maintenance costs. Cloud manufacturers, for their part, can operate through upstream-downstream synergetic and scalable processes, where these capabilities are deposited into open-box, tieredly integrated products. Depending on the stage at which the enterprise is located, it may be flexible to select the infrastructure, the Agent platform or the AI agent, which will be used to open the box, to quickly validate the idea and continue to overlap.
Baidu smart cloud Agent Infra covers sandboxes, computing, network, storage, observation and safety. Sandboxes are responsible for rapid start-up missions, isolation of operating environments and maintenance of mission status, allowing AI agent to carry out its tasks within secure borders; computing, network and storage, like “hydropower”, efficiently support mission execution, data access and context processing; and capable of documenting the AI agent implementation trajectory to assist enterprises in analysing costs, effects and causes of failure.
Through these capabilities, AI agent can operate in a permanent, stable and safe manner. On the ground, the Baidu smart cloud further lowered the threshold for AI agent transfers. In the past, complex interfaces and cumbersome configurations of cloud products have often impeded the automated implementation of AI agent; to that end, Baidu intelligent clouds have been tailored to Agent's call habits, and have further optimized the interface and usage of cloud products, allowing AI agent to move computing, storage and network resources on clouds more efficiently and directly, as in the case of “move local files”.
At present, Baidu smart clouds have continuously validated capabilities in their own products and client projects, providing complete support to enterprises from resource supply, AI agent operations to development operations.
III. Token demand surged and AI Infra needed to improve resource efficiency
After Agent Infra ran the task of AI agent, there was also a need for AI Infra to provide ongoing training, reasoning and resources. With the increase in the number and size of AI agent, the pressure on AI Infra is rising simultaneously. IDC predicts that the need for reasoning brought about by these AI agents will exceed 100 GVA. The more AI agent, the longer the mission chain, the higher the calculator, storage, network and energy inputs, the more cloud producers need to simultaneously expand their resources and increase the efficiency of Token output per unit of resources.
Zheng Zheng mentioned in the interview that the length of AI agentcontext is constantly increasing, that model parameters are constantly growing in size and structure, and that the share of post-training is also increasing. These changes place greater demands on computing, storage, network and reasoning efficiency, and infrastructure needs to be redesigned around model training and reasoning to enhance data access and token efficiency in a long context scenario.
The Baidu smart cloud supports the operation of AI agent more efficiently by means of a 100 舸AI computation platform 6.0, which provides a purely calculative platform and upgrades it to a “smart factory” capable of synergizing infrastructure, resource management and engineering capabilities.
AI agent went behind more scenes, requiring both Agent Infra to take over the tasks of AI agent, tool call and operations management, and AI Infra to continuously upgrade model training reasoning and resource efficiency. Together, AI agent will be able to maintain stable operation in a longer mission chain, larger calls and control overall usage costs.
Conclusion: Agent enters the production site and the new cloud base stabilizes demand
AI agent is moving from personal tools to implementation modules in the enterprise business system. Modelling capacity can be converted into sustainable reused industrial value only if it enters the production chain.
The current AI agent landing in a growing number of scenes is constantly generating new demands, and this “new” will also drive infrastructure industries such as chips, light modules, storage, etc., to new applications and business models.
For cloud manufacturers, the competition focus also extends from providing a single computing resource to a complete infrastructure that supports the long-term operation of AI agent. The collaboration between AI Infra and Agent Infra is becoming an important path to Agent's landing.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.