Baidu AI Cloud launches a dual-layer AI Infra and Agent Infra foundation to support the implementation of industrial agents

The scale of agents is expanding rapidly. IDC predicts that by 2030, the number of active agents worldwide will reach 2.216 billion, with a compound annual growth rate of 139%, more than doubling each year.

Rising in tandem with the hype is Token consumption. According to the National Data Administration, by March this year, China's average daily Token calls had exceeded 140 trillion, more than 1,000 times the level at the start of 2024; monitoring by the China Academy of Information and Communications Technology shows that by June this year, that figure had reached 500 trillion. Such a massive volume of calls shows that agents have moved beyond concept demos and begun to take on a large number of real tasks.

These frequently called agents are mostly concentrated in general capabilities and high-frequency scenarios. There is another route worth examining separately—industrial agents. General agents compete on scenario coverage and call frequency, whereas industrial agents plunge directly into business frontline settings such as community governance, factory floors, industrial parks and stores, solving trivial and complex real-world problems and daily affairs.

This is precisely a path that Baidu AI Cloud has forged. At the Baidu AI Cloud Super Agent Conference held in Wuhan on October 29, practical cases of Baidu agents being deployed emerged across various scenarios, from community governance and park operations to manufacturing and store management: they can automatically generate governance work orders based on residents' requests, turning public needs into traceable, closed-loop handling processes; they can also distill the experience of veteran factory workers into callable skills for quick reuse later.

Connecting these practices, it is not hard to see several characteristics of Baidu agents' deployment: first, the scenarios focus on the industrial frontline, with practical deployment cases in governance, operations, production, services and other links. Second, they directly confront complex needs, especially in experience-intensive, long-process fields such as industrial manufacturing, which place higher demands on agents' execution capabilities; third, they emphasize closed-loop tasks, enabling needs to produce results and experience to be reused.

For agents to run in enterprises over the long term and at scale, several challenges must first be solved: resource supply, task execution, data connectivity and security management. Baidu AI Cloud has therefore launched a two-tier infrastructure of AI Infra and Agent Infra, making agents a key productivity force for enterprises.

1. From personal assistants to core business drivers, agents enter the deep waters of industry

Over the past year, agents have been moving from personal efficiency tools toward core enterprise business scenarios, with deployment gradually advancing into deep waters. At the conference, Baidu Vice President Shi Qinghua said that Baidu Dazi, as a general agent, has reached tens of millions of monthly active users since launching in March this year, growing more than 10-fold month on month, and agents are rapidly penetrating the general public.

While becoming increasingly widespread on the consumer side, agents are also releasing industrial value in enterprise business scenarios.

After Hangzhou condiment company Songxianxian adopted Baidu Dazi, data aggregation work that originally took two days can be shortened to 3 minutes. The agent demonstrated cross-system data reading and business insight capabilities, beginning to connect enterprise internal systems and participate in business analysis and judgment.

Beyond enterprise business scenarios, Baidu Dazi is also deeply involved in grassroots governance. In Qiaokou, Wuhan, the grassroots governance project of government-enterprise service company Honghaixin based on Baidu Dazi has covered 1 million people, 300 users and 40 real scenarios. Information posted by residents in grid group chats can, after analysis by the agent, automatically generate request work orders and be pushed to the street office system for follow-up handling. In complex scenarios, Baidu agents possess unstructured data processing capabilities, automated execution capabilities and cross-system collaboration capabilities.

Whether for enterprise efficiency or grassroots governance, the deployment of agents is not achieved overnight; behind it is the reconstruction of an entire systems engineering effort. As agents' involvement in enterprises continues to increase, the related infrastructure also needs to be upgraded in tandem.

Zheng Ran, co-chair of the Baidu AI Cloud Technology Committee, said in an interview with Zhidongxi that enterprises are no longer just calling a large model API; to enable agents to go deep into business, cloud vendors need to provide two types of support: on one hand, out-of-the-box agent capabilities to help enterprises quickly integrate and experiment; on the other hand, infrastructure to support the long-term operation of agents, including runtime environments, data connectivity, security management and continuous operations and maintenance.

These two types of capabilities are precisely the underlying support that Baidu AI Cloud provides for enterprise agent deployment.

2. For agents to endure, after models, the "operating foundation" must first be shored up

For agents to deliver long-term value in enterprises, they need a secure, stable operating environment whose value can be measured.

First, agents need to call tools and enterprise internal data; a single operational error could cause data leaks or system crashes, so a secure execution environment is needed. A "sandbox" allows agents to execute tasks in an enclosed environment, avoiding impact on other systems.

Second, large models are stochastic, and enterprises also need to know what agents actually did, how much cost they incurred and how effective they were. This requires infrastructure with "observability" capabilities, helping enterprises see agent costs clearly, pinpoint problems and drive continuous iteration of results.

The essence of Agent Infra is precisely to turn one-off model calls into a task system that can run long-term, be observed and be governed. After an agent is deployed, the real difficulty is keeping it running stably in the enterprise environment and continuously delivering results. To achieve this, one must first understand the underlying technical form of agents. Mature cloud-native applications can start on demand, elastically adjust computing resources according to task volume, call cloud computing, storage and network resources in a standardized way, and have the platform continuously handle deployment, scheduling and operations and maintenance.

Zheng Ran told Zhidongxi that an agent itself is also a cloud-native application, and existing cloud-native technologies remain the foundation for its operation. However, agents' resource demands are more elastic, there is also a certain degree of uncertainty during operation, and in the future they may run at massive scale, so cloud-native infrastructure needs to evolve further to adapt to their workloads.

As the number of agents grows rapidly, infrastructure must also support larger-scale collaborative operation, which places higher demands on the underlying architecture. This means that after agents enter production, the core gap for enterprises gradually shifts from models to the runtime foundation.

Zheng Ran said that if enterprises build Agent Infra from scratch, they need to cover multiple technical domains such as computing, storage, networking, cloud-native and large-model engineering, and both the construction investment and ongoing maintenance costs are high. Cloud vendors, through upstream and downstream collaboration and large-scale operations, can consolidate these capabilities into out-of-the-box, layered and composable products. Enterprises can flexibly choose infrastructure, agent platforms or out-of-the-box agents according to their own stage, quickly validate ideas and continue iterating.

Baidu AI Cloud's Agent Infra covers capabilities including sandbox, computing, networking, storage, observability and security. The sandbox is responsible for quickly starting tasks, isolating runtime environments and saving task states, allowing agents to execute tasks within secure boundaries; computing, networking and storage, like "water and electricity," efficiently support task execution, data access and context processing; observability records agents' execution trajectories, helping enterprises analyze costs, effectiveness and causes of failures.

Through these capabilities, agents can run stably, securely and over the long term. On top of the underlying capabilities, Baidu AI Cloud has further lowered the threshold for agent invocation. In the past, the complex interfaces and cumbersome configurations of cloud products often hindered agents' automated execution; to this end, Baidu AI Cloud has deeply optimized the interactive interfaces and usage methods of cloud products specifically for agents' invocation habits, enabling agents to schedule computing, storage and network resources on the cloud more efficiently and directly, just like "calling local files."

At present, Baidu AI Cloud has continuously validated the relevant capabilities in its own products and customer projects, providing enterprises with complete support from resource supply and agent operation to development and operations.

III. Token demand explodes, and AI Infra needs to improve resource efficiency

After Agent Infra gets agents' tasks running, AI Infra is still needed behind the scenes to continuously provide model training, inference and resource supply. As the number of agents and the scale of invocations grow, the pressure on AI Infra is also rising in tandem. IDC predicts that the inference demand brought by these agents will exceed 100 gigawatts. The more agents there are and the longer the task chains, the higher the investment in computing power, storage, networking and energy, and cloud vendors need to simultaneously expand resource supply and improve the token output efficiency per unit of resource.

In the interview, Zheng Ran mentioned that agent context lengths are continuously increasing, model parameter scales continue to expand, structures are becoming more sparse, and the proportion of post-training is also rising. These changes place higher demands on computing, storage, networking and inference efficiency, and infrastructure also needs to be redesigned around model training and inference to improve data access and token efficiency in long-context scenarios.

Through the Baige AI Computing Platform 6.0, Baidu AI Cloud has upgraded the AI computing platform from simply providing computing power to an "intelligent factory" capable of coordinating infrastructure, resource management and engineering capabilities, more efficiently supporting agent operation.

Behind agents' entry into more scenarios, Agent Infra is needed to undertake agents' task execution, tool invocation and runtime management, while AI Infra is also needed to continuously improve model training and inference and resource supply efficiency. Only through the coordination of the two can agents maintain stable operation in longer task chains and larger-scale invocations while controlling overall usage costs.

Conclusion: Agents enter the production floor, and a new cloud foundation stabilizes implementation needs

Agents are transforming from personal tools into execution units in enterprise business systems. Only when model capabilities enter the production stage can they be converted into industrial value that can be continuously reused.

At present, agents are being implemented in more and more scenarios, continuously generating new demands, and this "newness" will also drive the development of infrastructure industries such as chips, optical modules and storage, and spawn new application scenarios and business models.

For cloud vendors, the focus of competition has also expanded from providing a single computing resource to complete infrastructure that supports the long-term operation of agents. The coordination of AI Infra and Agent Infra is becoming an important path for agent implementation.