TERMINUS
With the wave of AIGC (Generative Artificial Intelligence) sweeping the globe, the evolution of technology is moving from simple content generation to a deeper “AI Agent” revolution. This is not only a technological iteration but also a disruptive revolution in productivity paradigms. Against this backdrop, at the 2025 World Artificial Intelligence Conference (WAIC 2025), the well-known technology think tank ZhiZi GuangNian released the “Value and Application Report of Enterprise-level AI Agents”, focusing on the commercial value and implementation paths of enterprise-level intelligent agents. Teslion, with its space intelligent agents designed for B-end scenarios and the HALI intelligent agent system launched for the consumer market, is selected as a typical case in the report.
Image source: “Report”
AI is entering a new stage, with “tool invocation” becoming a core feature of intelligent agents.
2025 is often seen as the “Year of the Agent”; however, discussions surrounding AI Agents are still filled with confusion: What exactly is the concept of an Agent? What factors contribute to the concentrated emergence of Agents in 2025? What capabilities constitute the core features of an Agent? According to ZhiZi GuangNian, the rise of AI Agents is not accidental; the synchronous development of large models, computing power supply, energy supply, open source, ecosystems, and industrial applications has collectively “lifted” the timely birth of AI Agents, making them one of the hottest technological trends today.
Referring to OpenAI’s five-level classification of AI, the report points out that at the current point, AI has progressed from L1 (chatbots) to L2 (reasoners), and is developing towards L3 (agents), with a push towards L4 (innovators) and L5 (organizers).
Driven by industry leaders, major companies like OpenAI, Anthropic, and Google have successively released key Agent products and technical protocols, with relatively mature products such as Manus, OpenAI Operator, and Terminus HALI beginning to emerge, making intelligent agents gradually become productivity tools for enterprises. However, the explosive popularity of the concept has also exacerbated the confusion surrounding the definition of intelligent agents—there are differences in definitions between academia and industry, and even among industry giants, the details of their definitions vary.
Integrating multiple definitions, the report suggests that intelligent agents can be viewed as systems that integrate large model capabilities and possess automated tool attributes; for enterprise-level intelligent agents, “tool invocation” constitutes the most core feature of the agent.
Image source: “Report”
Following this line of thought, the report delineates six capability dimensions required for intelligent agents—dialogue ability, reasoning ability, memory ability (specifically long-term memory), tool ability, planning ability, and action ability. Among these, dialogue ability and reasoning ability are core capability requirements for large language models/chatbots; based on these, Copilot (intelligent office assistant) must additionally possess long-term memory ability. Tool ability, planning ability, and action ability are the core indicators that distinguish intelligent agents from Copilots. The report states that in the Agent scenario, AI is capable of completing the vast majority of tasks, while humans are responsible for setting goals, providing resources, and supervising results; AI handles task decomposition, tool selection, progress control, and autonomously ends work upon achieving goals.
The digital economy accelerates the development of intelligent agents: Teslion’s four types of Agent roles each play their part.
The report points out that China’s level of digitalization and informatization continues to grow, providing opportunities for the digital economy and enterprise-level AI, and accelerating the application of Agents in enterprise scenarios. Through interview analysis, the report categorizes the application of intelligent agents in enterprise scenarios into four roles: “the brain consultant for front-end interaction”, “the inspiring partner for thoughtful feedback”, “expert-level digital employees”, and “creators that bring imagination to life”.
Image source: “Report”
Taking Teslion’s intelligent agent product system as an example, the large model solution for the real estate sector developed in collaboration with Colliers acts as the “brain consultant for front-end interaction”. This solution leverages Colliers’ diverse professional service experience in vertical fields and Teslion’s R&D capabilities at the infra layer, model layer, and agent layer, providing services such as industry consulting, corporate encyclopedia, and generative reports, thereby enhancing the experience and intelligence level of customer service with an “external brain” to solve single-point problems using “collective wisdom”.
Additionally, the “Healthcare AI Convenience Assistant” developed by Teslion assumes the role of an “expert-level digital employee”. This product achieves a fully domestic configuration from computing power and models to applications, based on the prototype of healthcare system staff in Zhengzhou, integrating the “healthcare voice assistant” with the Deepseek large model to provide citizens with 24/7, real-time, intelligent healthcare consultation and business processing. Currently, the related product has been fully launched in the Zhengzhou healthcare system, empowering healthcare services through AI technology.
The enterprise office assistant AI is the “creator that brings imagination to life”, leveraging the large-scale efficient generation capabilities of AIGC to achieve low-cost, high-quality creative material generation, accelerating the content production process.
The first BUTTONS CLIP headphones equipped with the Teslion HALI intelligent agent system
The HALI intelligent agent system serves as both the “thoughtful feedback inspiring partner” and the “all-around assistant with strong planning capabilities”, providing personalized feedback to users while becoming a caring assistant that helps users manage everything…
As the upstream and downstream of the intelligent agent industry chain gradually become complete, and as intelligent agents penetrate deeper into vertical industries, a comprehensive ecological map of enterprise-level intelligent agents in China covering the infrastructure layer, platform layer, and application layer is gradually taking shape.
Image source: “Report”
Two Teslion products selected, achieving breakthroughs in both consumer and enterprise levels.
As a pioneer and leader in China’s AIoT industry, Teslion is committed to using technology to drive industrial upgrades and spatial intelligence for sustainable development. Focusing on spatial intelligence, Teslion concentrates on three strategic directions: AIoT infrastructure, AIoT domain models, and AIoT intelligent agents, building a one-stop solution that integrates domestic computing power, domestic models, and domestic operating systems, creating a Chinese technology brand for the international stage.
Intelligent agents are the comprehensive interactive entry point of Teslion’s complete spatial intelligence solution. Relying on strong underlying technical capabilities, Teslion’s self-developed intelligent agents possess four capabilities: human-like thinking, long-term memory, team collaboration, and high-dimensional perception, using robots and smart wearable devices as carriers to bridge the last mile of AI large-scale applications.
The report includes Teslion’s spatial intelligent agents designed for spatial intelligence scenarios and the HALI intelligent agent system aimed at C-end application scenarios as typical cases of intelligent agent products. Among them, in spatial intelligence scenarios, spatial intelligent agents are the main executors interacting with human users through language and behavior, capable of understanding and processing various spatial-related tasks, and can provide spatial execution and services for other intelligent agents, enriching user experience in spatial contexts; as a companion intelligent agent, HALI boasts over 97% semantic understanding accuracy, with a text chat response delay of about 500 to 1500 milliseconds based on a vast knowledge base, and in translation mode, HALI achieves a bidirectional translation accuracy of 96% between Chinese and English.
Image source: “Report”
The report specifically mentions the core technological breakthroughs achieved by Teslion’s intelligent agent series products.
▶ Using end-to-end reinforcement learning to overcome the “human orchestration” dilemma of intelligent agents: Employing an end-to-end reinforcement learning method, the HALI intelligent agent system is fine-tuned with high-quality data, optimizing strategies through a reward function rather than relying on manual input, guiding the model to continuously improve through interactions with tools/devices/environments.
▶ Using knowledge graph-based data compression storage and retrieval technology to alleviate the cost and efficiency pressures of long-term memory: Introducing efficient data compression methods to extract key effective information from user dialogue data, filtering out unimportant parts, thereby increasing the information density of stored data and reducing storage volume. On this basis, HALI constructs a user knowledge graph to enhance the system’s understanding of user intent and the accuracy of retrieval results during the retrieval phase.
▶ A hybrid solution of small models on mobile devices and large models on servers to address multi-agent collaboration latency challenges: Utilizing parallel multi-model execution technology, employing a hybrid solution of fine-tuned small models on mobile devices and large models on servers, ensuring low latency in user control command responses and accuracy of control commands.
In the future, Teslion will further balance breakthroughs in underlying model capabilities and innovations in product engineering, connecting more perception terminals and vertical scenarios, accumulating higher quality and more dimensional data, and building intelligent agent products capable of understanding richer meanings of the physical world.
Further Reading
If there are any copyright issues with the materials used on this platform, please contact the author.