Introduction: Sober Reflection Amidst Technological Hype

From the explosion of large language models to the gradual popularization of edge computing, from intelligent voice assistants entering homes to smart devices connecting to the cloud, the combination of AI agents and AIoT (Artificial Intelligence of Things) is becoming a new hot topic in the industry. However, behind this wave of technological enthusiasm, we must calmly consider a core question: Are we truly improving the status quo, or are we once again falling into an "intelligent illusion"?

Currently, the concept of AI agents carries the risk of being over-packaged. As Gartner pointed out in its report, many so-called AI agent systems are merely "conversational robots endowed with a sense of task," lacking true perception, action, and task closure capabilities. This phenomenon is not new; a 2018 Cisco research report showed that 75% of IoT projects globally ultimately failed. The fundamental reason was a lack of clear goals and intelligent control logic, leading to a blind pursuit of "connection for connection's sake," which ultimately increased complexity rather than wisdom. Therefore, the problem lies not with AI or IoT themselves, but with "castle in the air" development paths that deviate from task closure and scenario value.

This article aims to re-examine this issue and emphasize that the true path for AI agents lies in deep coupling with real physical systems—namely, AIoT. The Internet of Things serves as the foundation for AI to connect with the physical world. Only when AI agents are deeply integrated with IoT devices and embedded in the physical world can they escape the fate of "intelligence showing off" and truly become value creators in AIoT scenarios. What we seek is not just smarter tools, but a wiser system.

This article will delve into three aspects:

1. The true value and system structure of the combination of AI agents and AIoT.
2. How AI agents can break through the bottleneck of "demonstration intelligence" and move towards "scenario intelligence."
3. How to validate the boundaries of agent capabilities through "real task closure" to avoid the next technological bubble.

Why AI Agents Must "Land"? From Virtual Intelligence to Physical Closure

AI agents currently face a fundamental bottleneck: most of them grow in virtual environments, lacking interaction with the real world. This detachment from the physical world limits their ability to continuously learn and optimize. Only when AI agents begin to interact with real devices, receive real sensor data, perform physical actions, face environmental uncertainties, and take responsibility for the results, can they truly possess the potential for continuous learning and optimization. This evolution from closed input-output to task feedback loops is a critical turning point for intelligent systems, transforming them from passive tools into active agents. All of this can only happen within an AIoT system.

The essence of AIoT goes far beyond simple device networking; it emphasizes the "taskification" of systems. Within the AIoT framework, every connected device and every sensor node becomes part of how AI agents understand and influence the world. Rather than viewing AIoT as a purely technical domain, it is better understood as the "task environment" for AI agents. In this environment, AI agents are no longer passive APIs waiting to be called, but autonomous executors with a clear sense of purpose, scheduling authority, and feedback mechanisms.

Taking smart warehousing systems as an example, AI agents not only need to understand order structures and inventory statuses but also must dynamically schedule multiple robots, plan paths based on ground conditions, traffic density, and task priorities, and continuously adjust strategies during execution. Similarly, in automated charging station dispatch, AI agents must predict future load peaks, identify vehicle types, determine battery status, and make optimal allocations based on real-time grid loads. These are concrete examples of AI agents "landing" in the real physical world.

Of course, AIoT itself is also continuously improving. The fundamental reason why traditional IoT projects often fall into the trap of "connected but not intelligent" is the lack of an intelligent entity that understands "why connect." In this context, the introduction of AI agents is not merely an embellishment but a structural upgrade. AI agents can not only interface with sensory inputs but also integrate business rules, objective constraints, and system capabilities to form behavioral outputs with automatic planning and adaptation capabilities. Therefore, only when AI agents are deeply embedded in AIoT systems and truly participate in the perception and action of the physical world can they possess a real task environment and feedback mechanism, thereby creating true value.

From "Model Access" to "Scenario Intelligence": AIoT Needs Systems, Not Plugins

The true value of AI agents is not merely in calling an API and returning an answer, but in their ability to handle entire tasks and serve as responsible system roles. This means that intelligent systems are not simply about grafting models, but about re-architecting system capabilities.

In real-world AIoT systems, a single AI agent often cannot handle all tasks. A highly dynamic and task-variable system (such as a smart factory, smart building, or urban energy system) does not require "one super-intelligence that governs everything," but rather a systemic intelligent structure where multiple agents collaborate with clear responsibilities. For example, in a smart factory, the scheduling system needs to set task rhythms based on order and inventory status, the quality inspection system needs to determine if products meet standards, the logistics system needs to arrange finished product outbound paths, and the maintenance system needs to monitor equipment health and schedule repair windows. Each subsystem has its own task environment, data interfaces, and feedback mechanisms. Attempting to use one large model to handle all problems would not only be inefficient but could also lead to systemic risks due to confused responsibilities.

The truly effective architecture should involve multiple specialized agents collaborating through shared perception, limited communication, and clear boundaries. This multi-agent collaboration model is not only more aligned with engineering practices for maintainability and scalability but also better reflects the operating principles of complex systems. In this integration, AI agents are responsible for cognition and decision-making, while IoT devices handle perception and execution. Throughout this process, agents are not only task decision-makers but also responders to environmental changes and coordinators of system resources.

In contrast, some "pseudo-intelligent" cases often devolve into superficial interaction upgrades: a console adds a voice assistant, a home system gains a chat interface. While seemingly enhancing the "smart experience," the core capabilities remain unchanged—the system does not truly understand task objectives, devices remain isolated, and user intentions are not translated into system collaboration. This "new wine in old bottles" approach not only fails to deliver substantial value but can also exacerbate user misunderstanding and fatigue towards AIoT.

Worth Investing, Yet Prone to Misdirection: AI Agents + AIoT at a Crossroads

Currently, the convergence of AI agents and AIoT stands at a critical juncture—one that is both highly promising for investment and highly susceptible to misdirection. On one hand, the technological foundations are maturing, seemingly setting the stage for widespread adoption. On the other hand, the landscape is rife with misunderstandings, oversimplifications, and hype, making the realization of true value more complex and challenging.

From a technological evolution perspective, the combination of AI agents and AIoT is entering an unprecedented window of opportunity. Firstly, the standardization of IoT devices has significantly improved, with communication protocols, edge computing frameworks, and data interfaces gradually unifying. This greatly lowers the barrier to deploying and integrating agents. Secondly, advancements in AI model fine-tuning and reinforcement learning technologies have enabled agents to learn from tasks and optimize based on feedback, marking their transition from research prototypes to deployable systems.

However, precisely because the technology appears "usable," the risks become more subtle and alluring.

One common misconception is mistaking "accessing a large model" for "possessing an intelligent agent." The danger of this illusion lies in its ability to create a "sense of intelligence" without any systemic intelligent capabilities. Once introduced into complex environments, it exposes fundamental flaws such as chaotic decision-making, uncontrolled execution, and a lack of accountability.

Another pitfall is neglecting the execution loop of tasks. A truly effective agent must possess the ability to track task status, plan and adjust execution paths, and implement mechanisms for result verification and feedback.

A more fundamental issue is the lack of scenario design capabilities. An effective AIoT system is not merely a collection of technologies but an intelligent task network built around specific scenarios. This requires system designers to understand both technology and business processes, enabling them to embed the "sense-understand-act-feedback" loop into real user journeys.

Therefore, the integration of AI agents and AIoT is a path worth pursuing, but it is by no means a "shortcut."

How to Make "AI Agents + AIoT" Move Beyond Illusion and Take Root in Reality?

As the convergence of AI agents and AIoT gradually becomes the new wave of technological enthusiasm, how to prevent this trend from repeating the mistakes of past "tech bubbles" becomes an unavoidable question. History has repeatedly shown that technology itself does not fail; rather, it is often unrealistic expectations, disconnected scenario designs, and poorly governed system structures that lead to failure.

A truly intelligent system is not built for demonstration but for continuous operation. This means that from the initial design phase, deployment must be considered, including task lifecycle management, optimization of resource scheduling, handling of abnormal states, and long-term interaction between users and the system. In this process, a systematic evaluation framework is indispensable. For example, Salesforce launched CRMArena-Pro, the first multi-turn enterprise-grade benchmark test for agents, providing a comprehensive evaluation system for task completion rates, multi-turn interaction capabilities, policy compliance, and system security. This allows developers to identify design flaws early on and prevent products that "look smart" from collapsing in real-world environments.

From the perspective of corporate strategy and development practice, we also need a set of criteria to identify which "AI agents + AIoT" projects have real value and which are merely illusions packaged by marketing. The following four questions can serve as preliminary judgment criteria:

1. Does the system have a complete "sense-decide-act-feedback" loop? This is the foundation for the effective operation of an intelligent system, ensuring it can acquire information from the environment, make judgments, take actions, and adjust based on results.
2. Does it solve an existing efficiency bottleneck or human pain point in the real world? True value lies in solving practical problems, improving efficiency, or compensating for labor shortages.
3. Can the system's operation be quantitatively evaluated using metrics such as ROI (Return on Investment), task completion rate, and error rate? Quantifiable metrics are key to measuring the value and effectiveness of a system.
4. Does the agent participate in task execution, or is it merely an interface or data query tool? Agents should be active executors, not passive tools.

True intelligence is not marked by peak performance in solving top-tier problems, but by the continuous, stable, and error-free generalization ability across all simple and complex tasks.

In some industries, the value of AI agents has already begun to emerge. In the field of precision medicine, French AI biotechnology company Owkin has built a cancer research AI agent that integrates multimodal data from over a million patients, promoting target identification, patient classification, and clinical trial optimization, significantly improving personalized treatment. In infrastructure and public services, agents are also driving the intelligent reconstruction of traditional systems. For example, the "Water Radish® AI Agent" independently developed by Jinkehuanjing has been deployed in 5 water plants in the Wuxi area, achieving L4 unmanned operation, replacing over 90% of manual daily tasks, reducing operation and maintenance teams by 60-70%, energy consumption by 15%, and overall operating costs by 35%. It has also passed the authoritative evaluation of the Ministry of Industry and Information Technology's "AI Industry Innovation Scenario Application Case."

Conclusion

AI agents are not a panacea, nor is AIoT a universal container. The true value of AI agents lies not in "making devices smarter," but in building a more collaborative system structure. The combination of AI agents and AIoT is a critical turning point that propels AIoT from "connection" to "intelligent collaboration," but it is by no means the end of technological evolution.

Indeed, if we merely deploy a large model in every device and an agent in every system just to "appear smarter," we will ultimately create "pseudo-intelligent systems" that are uncoordinated and difficult to maintain. What is truly worth pursuing is not that every device can converse, or every terminal can reason, but that the entire system can form a dynamic, efficient, and controllable collaborative network around real tasks. Future AI innovations may not have stunning interfaces, fluent semantic generation, or even be particularly "human-like," but they will be able to genuinely take responsibility, solve problems, and create value.

References

1. CRMAArena-Pro: Holistic Assessment of LLM Agents. Across Diverse Business Scenarios and Interactions, Source: Salesforce AI Research
2. Agentic AI: the evolution of intelligent automation, Authors: Paras Sharma, Joydeep Bhattacharyya, Source: Transforma Insights
3. Salesforce AI Launches CRMArena-Pro: First Multi-Turn Enterprise Benchmark for LLM Agents, Source: nxrte.com