Agent is a word that cannot be avoided in 2026. The direction itself holds—having AI not only give advice, but complete a sequence of actions on its own within its permissions, does have value for process-heavy work such as operations.
But "the direction holds" and "it can be fully deployed now" are two different things. What sits between them is not model capability. It is infrastructure capability: for an agent to act on its own on an industrial site, the prerequisite is that it can sense the site accurately, and that its actions have a boundary and a record. Both of those sit at the digital IoT layer. This is also the order the first nine articles in this series have been stating—the ceiling of industrial-grade AIoT is set by what kind of data it can obtain.
Two bottlenecks the industry itself has pointed out
According to public reports, Wu Hequan, an academician of the Chinese Academy of Engineering, said in his keynote at the 2026 Industrial Internet Conference that industry is entering a new stage driven by agents. In the same speech, he also pointed out two bottlenecks that are not yet solved: on-site industrial control equipment (PLCs and industrial PCs) does not have enough native compute in reserve, which constrains the interconnection of cross-vendor agents; and there is a foundational conflict between the millisecond-level determinism required by industrial control and the inherent randomness of large-model decisions, with no mature general solution at present.
Read those two sentences together with the sentence before them, and the conclusion is clear: the judgment of the stage is optimistic, and the conditions for deployment are strict. On-site sensing, local processing, and deterministic execution are the prerequisites for an agent to actually do work in an industrial scenario. They are not accessories to be added after it is deployed.
Deployment at the policy level points to the same place. The Action Plan for Promoting Innovative Development of the IoT Industry (2026—2028), issued in March 2026 by nine departments including the Ministry of Industry and Information Technology, lists as priority directions the coordinated deployment of sensing intelligence and large models on terminals and the edge, and unified data standards—terminals and the edge are exactly where sensing and execution happen.
A practical investment order
The suggestion for decision-makers is four steps. Spend money in this order:
Step 1: connect the equipment that should be connected, and keep the data trustworthy. This is not only access. It also means solving offline caching, time synchronization, unified definitions, and ledger context. The standalone return of this step is that status is visible and an anomaly does not have to wait for the customer to call.
Step 2: pick one scenario, run the closed loop through, and keep records. Alarm grading, responsibility assigned to a person, work-order dispatch, handling feedback, and report statistics. The standalone return is that responsibility is clear, handling can be traced, and the monthly report does not have to be assembled by hand. At the same time it is producing labeled data for what comes later.
Step 3: bring AI assistance into steps that are verifiable and reversible. Alarm summaries, similar-case retrieval, draft reports, and risk prompts, each with human review and effect metrics. The standalone return is that repeated information-organizing work decreases.
Step 4: then evaluate agents. By this step you have a clean data pipeline, closed-loop records, and data on the actual effect of AI within a limited scope. Judging then whether agents are worth the investment has a sufficient basis.
The most important benefit of this order
Each step has value on its own. There is no need to bet that only the last step pays back.
This is worth saying on its own, because it sets the risk structure of the project. If the whole investment is pressed onto the endpoint of "building a smart factory," and any step in the middle does not go smoothly, the money already spent is hard to show as value, and the project is easy to cut when leadership changes or the budget tightens.
Follow the order above, and finishing step 1 brings an improvement in operations efficiency, finishing step 2 brings an improvement in the basis for management, and finishing step 3 brings an improvement in labor input. Whether to do step 4, and when, can be decided from the actual data of the first three steps. When the technical route changes, the earlier investment is not written off—the data pipeline is a general-purpose asset, and it is not bound to one generation of models.
Where Tongwei Smart sits on this path
Stating clearly what we do and what we do not do is more useful than stating it in larger terms.
Guangzhou Tongwei Smart Technology Co., Ltd. develops smart IoT gateways, expansion modules, industry terminals, and the AIoT platform in-house. What it does is the first two steps of this path, and the verifiable part of step 3. The business has two main lines. Standard product pairing pairs mature smart IoT gateways, modules, industry terminals, and interfaces to equipment manufacturers, control-cabinet plants, ELV contractors, and integrators, and answers where the data comes from. Solution pairing packages hardware, the platform, implementation, and operations by scenario, and answers how one site is gotten through as a whole.
The two lines share the same device–edge–cloud base. It has already been deployed in these six scenario types: equipment-room power and environment, smart public toilets, building automation, energy-efficiency management, smart agriculture and aquaculture, and urban ecology and public facilities. What changes is the points and the rules, not a rebuild of the system.
We do not build models, we do not promise to build a smart factory for the customer, and we do not promise a timetable for agents to be deployed. Tongwei Smart's place is to turn scattered equipment, environment, and energy sites into objects that can be connected, diagnosed, and operated—so that later AI has something it can use. This place does not need to be raised. It is itself the link most lacking at present.
The boundary is unchanged: we do not replace the customer's core controllers or on-site safety systems; whether a specific connection is possible is assessed from the equipment list, protocols, and point tables; at this stage the focus delivered is visual monitoring, active alarms, and an operations closed loop, plus verifiable assisted diagnosis and reports, and predictive maintenance is opened gradually as data accumulates.
Next step
If you are making next year's digital budget, you can first do something that costs very little: take stock of the current equipment count, protocol types, share of equipment that is connected, how anomalies are found, and how handling is recorded. That stocktake tells you directly which of the four steps you stand on now, and where the first sum of money should be spent.
If needed, Tongwei Smart can join this stocktake, or do a requirements diagnosis directly.