Technology is surging ahead, demand is emerging, old and new are alternating, and the endgame has yet to be seen. The industry is both excited and anxious.

In the first half of 2026, the implementation of AI agents has caused dramatic changes across the industry. “The pace of technological progress far exceeds imagination,” said Zhi Zhen, Chairman of China Industrial Internet, describing himself as being in an “exhilarated state,” dedicating 99% of his energy to technology and working on-site with clients to drive the integration of scenarios and technology. “As long as we can make a good product, there will definitely be buyers.”

However, the rapid advancement of the technological wave has also triggered widespread anxiety among enterprises. A source from a major company revealed that nearly all enterprises they met this year were asking: “The technological architecture is changing too fast. We’ve already invested in building an AI agent system, but now OpenClaw has emerged. Should we scrap it and rebuild? Will the AI agent projects we did last year become obsolete?”

The coexistence of technological surges and decision-making confusion defines this phase of AI agent implementation. How can the conflict between old and new architectures be resolved? What fundamental shifts are occurring in customer needs? And as AI agents move from edge pilots to core businesses, what path will the next phase of competition take?

01 The Old Map is Now Obsolete

The product design and philosophy of OpenClaw have brought many insights to the industry, as well as a shockwave, necessitating a complete rewrite of development efficiency, product forms, personnel structures, and software logic.

AI application capabilities have become the core evaluation criterion. “Our sole criterion for evaluating people now is their ability to use large models,” said Zhi Zhen. “One person who uses it well can outperform two or three who don’t.” Yu Youping, President of Zhongguancun Kejin, stated that in March this year, the company launched the OpenClaw Innovation Competition internally, where employees from different departments formed teams to directly build “Lobster Applications.” They received 95 good ideas corresponding to specific business actions, some of which have already been promoted and implemented internally.

Product forms are also undergoing dramatic changes. Yu Youping believes that OpenClaw has made the industry realize that “enterprise-level AI agent architectures can no longer be designed around ‘chat windows’ but should be designed around ‘work tasks.’” In the past, many AI agent products were essentially “large models + knowledge bases + dialogue interfaces,” making them ill-suited for enterprise-level tasks. What enterprises truly need is a work system capable of understanding tasks, invoking tools, using knowledge, adhering to permissions, outputting results, and accepting evaluations. The construction of AI agents is moving toward business personnel, allowing frontline staff to precipitate (translate: codify) their experience into Skills, with the delivery target becoming “digital employees.”

The market has responded quickly, with major companies like Tencent, ByteDance, Alibaba, and Baidu, as well as first-tier AI agent implementation vendors, all launching their own “Lobster-like” products. “Previously, we might have only been able to handle simple approval scenarios, but now we can handle autonomous operation scenarios, with more opportunities for new scenarios,” said Wang Zhong, co-founder of Zhongshu Xinke.

Parallel to the architectural transformation is a leap in development efficiency. In December last year, significant breakthroughs were achieved in large model programming capabilities. “With the same number of people, we can now handle at least twice as many projects simultaneously,” said Wang Zhong. Now, Proof of Concepts (PoCs) are rare. “We start by showing a demo to help clients understand. Then, we either let them try the product directly or discuss project delivery.” Building a complete, operational scenario demo “can basically be done within a day,” representing a huge boost in productivity.

Zhi Zhen cited an example: previously, creating a Bill of Materials (BOM) required selecting items one by one from dropdown menus. Now, with a small Claw tool, even if you only remember the name of a part, after inputting it, the system can automatically traverse the database, think and verify, and return a standard Excel sheet. “This functionality wasn’t entirely impossible last year, but it required vector libraries, fuzzy matching, and a bunch of other technologies, and the results weren’t great even after a lot of effort. This year, it’s simple—Lobster acts as a scheduling shell, and you just need to make a simple skill to connect to the database.”

The level of intelligence is also soaring. “May is at least 10 times better than January,” said Zhi Zhen. Previously, creating an AI agent that could recognize material codes from images or text and generate error-free purchase orders took months of development and couldn’t achieve industrial-grade accuracy. Before January, it was almost just a PoC, difficult to truly implement. Now, as long as you can think of it, you can build it, and the agent can repeatedly search, verify, and review, achieving multi-agent collaboration and significantly improving accuracy.

AI agents are also reshaping traditional software paradigms. In the past, software underwent “annual” iterations; now, “daily disposable software” has become a reality. Yu Youping noted, “The reshaping of traditional software paradigms is happening faster than imagined.” Enterprises are finding that AI can complete tasks across systems directly through “intention-driven” approaches, causing their patience with traditional SaaS to plummet. This shift has exploded in a short few months.

In the industrial sector, “industrial design simulation has undergone significant disruption this time around,” pointed out Zhi Zhen from China Industrial Internet. CAX software, such as CAD and CAE, once known as the “jewels in the crown of industry,” are now facing disruptive changes. The solid barriers built over time through code accumulation and vast model libraries are being broken by AI. “In January, I said industrial software might have a 3–5-year barrier, but now I think there’s none at all.”

Zhi Zhen predicts that the arrival of industrial AI agents is now just a matter of “when they will be fully rolled out,” with a time window of only 1–2 years. “First, there will be no such thing as industrial software this year—all enterprises undergoing informatization will add AI agents.” Previously, factories had to purchase a dozen or twenty software suites, with 99% of the functions in each suite possibly never being used. Now, they only need to generate lightweight algorithms tailored to specific scenarios, which are more effective, cost less, and can highly integrate CAD, CAE, MES, PDM, etc. In the future, software will no longer be about comprehensive functionality but will be based on demand and application scenarios, delivering value and functionality.

02 The Debate Between Old and New Architectures: Must We Start Over?

The rapid iteration of technological waves has made the issue of “old versus new” a focal point of anxiety for enterprises. Many companies are caught in a dilemma: should they scrap the AI agent systems they’ve already invested in and rebuild them with new frameworks like OpenClaw? Will previously implemented projects be quickly rendered obsolete?

Yu Youping from Zhongguancun Kejin believes that “the core issue for enterprises is not whether their old systems have value but whether the original architecture is open and flexible enough.” The difference between old and new systems lies in the fact that the new paradigm emphasizes two directions: enterprise-level AI agents should be organized around tasks and job-specific Skills, and the agent system should possess greater openness, scalability, and continuous operational capabilities. If the system built by enterprises last year adopted an open architecture, with decoupled capabilities such as models, tools, and knowledge bases, the cost of switching to accommodate new Skills systems, integrate new tool protocols, and upgrade to new frameworks would be relatively controllable, with many capabilities continuing to evolve on the original foundation. Conversely, if the construction was closed, tightly bound, and siloed, they would feel significant architectural pressure this year.

“What enterprises care most about now is not chasing the latest framework but avoiding being locked into a technological path. Because AI agents are still evolving rapidly, enterprise-level architectures must allow for sufficient flexibility,” said Yu Youping.

Wang Zhong from Zhongshu Xinke offered another layer of insight: clients don’t care about the paradigm used, only the results. As long as a project can be implemented and pass acceptance within defined scenarios, with satisfactory results, clients won’t demand that old scenarios be retrofitted with new paradigms. They would rather use new paradigms to tackle new scenarios that were previously unfeasible.

“After OpenClaw gained popularity, many existing clients approached us to explore new scenarios,” said Wang Zhong, giving the example of a leading company in the water industry that inquired this year whether the “Lobster” paradigm could handle scenarios that were previously impossible. They had already successfully implemented AI agent applications and are now advancing two new scenarios—process scheduling and leakage warning—based on the new paradigm.

In reality, since the rise of the Agent concept in 2023, there have always been two paths in the industry: one is workflow-based, and the other is more autonomous Agent-based planning. OpenClaw and Hermes, which emerged this year, are typical representatives of the latter path.

“Most of the platforms companies originally bought were workflow platforms, but now the intelligence and generalization capabilities of multi-agent collaboration far exceed those of workflow platforms,” said Wang Zhong. After demonstrating product prototypes featuring multi-agent collaboration to clients, they all preferred to use this paradigm as the foundation.

However, old and new technological architectures are not entirely substitutive but rather suited to different scenarios.

Within enterprises, there are numerous scenarios with fixed rules, rigorous processes, and highly certain outcomes where workflows already provide excellent solutions. For example, contract reviews and copywriting generation involve relatively fixed processes and knowledge, making workflows the best and fastest choice. The self-planning route would consume large amounts of additional tokens, making it inefficient and costly. Simple data and trend analyses also suit workflows.

In contrast, the new paradigm of multi-agent collaboration and self-planning, represented by OpenClaw, is suitable for scenarios with stronger generalization where processes cannot be exhaustively listed. Its greatest advantage lies in openness, complementing workflows.

Wang Zhong cited predictive equipment maintenance as an example. For diagnosing already occurred faults, where processes and knowledge are relatively fixed, the workflow mode suffices. However, predicting higher-value hidden risks, which cannot be exhaustively defined by abnormal rules and were previously difficult to address, can now be resolved through multi-agent frameworks that simulate human expert experience for reasoning and analysis. Accuracy has now exceeded 90%, and the system possesses continuous learning capabilities.

Nevertheless, the large-scale implementation of the new paradigm still faces many challenges. Wang Zhong admitted that while enterprise clients recognize its value, they are generally concerned about secure implementation. “The growth of AI agents is evident, but only a small proportion of scenario constructions dare to use self-planning structures. There is a lack of trust in the underlying framework, which still needs validation. Widespread adoption depends on the emergence of an enterprise-level framework.” He predicts that OpenClaw is not the final technological solution and that “new frameworks will continue to emerge,” with AI agent technology iterating continuously to evolve mature solutions for complex enterprise scenarios.

03 What New Demands Are Emerging from Clients?

In the first half of 2026, the demand for AI agent implementation has significantly accelerated. Zhi Zhen observed that the client base is becoming increasingly diverse, with no enterprise doubting the value of AI anymore—they just want to know “how much they want to spend to achieve what.” Demand patterns have also clearly upgraded: “Previously, clients said they didn’t want copilots anymore and wanted AI agents to take the lead, not humans.”

The enthusiasm for procurement among business departments has surged, even surpassing that of IT departments. Xin Zhou, General Manager of Baidu Smart Cloud’s AI and Large Model Platform, revealed that among the multiple clients they serve in industries like ports and manufacturing, business departments are now actively raising requirements and even directly procuring solutions. An industry insider also confirmed, “If the business department says it’s good, we proceed; if not, we don’t bother.” Business experience has become the core decision-making criterion.

Procurement logic is also changing: while clients recognize the value of platform foundations, they no longer pay for bare platforms. Wang Zhong from Zhongshu Xinke bluntly stated, “Selling basic platforms detached from scenarios is no longer feasible. There must be a strong core business scenario as a guide to incorporate these platforms and infrastructure as necessary modules.” This shift is directly reflected in budget structures. Taking several leading enterprises Wang Zhong serves as examples, the proportion of procurement for platforms and computing infrastructure has significantly declined, with resources rapidly shifting toward scenario-based AI agents. “Last year, it might have been a 60-40 split; now, it’s probably reversed to 40-60, with at least a 50% increase in AI agent orders.” The market now either elastically scales computing resources as needed or directly builds dedicated scenario-based AI agents; few clients procure general-purpose platforms like commercial versions of Dify or HiAgent in advance and then plan scenario implementations separately.

“Nowadays, everyone talks about scenarios and calculates costs very precisely,” said Wang Zhong. “Don’t talk to me about vague capability-building; I want to know where I can use this. Among the options 1, 2, 3, 4, 5, 6, prioritize those with the highest ROI.” ROI calculation methods are also deepening. In addition to traditional metrics like accuracy and expert agreement rates, some leading central state-owned enterprises have introduced Full-Time Equivalent (FTE) measurement methods this year, using core manual work hour substitution rates as key parameters.

At the same time, clients place greater emphasis on comprehensive enterprise-level functionalities. “Today, if you show me a demo and it’s of good quality, I won’t let it go live immediately. I’ll definitely ask how knowledge permissions are managed, how multiple verifications are done, what modifications are needed to integrate with existing business processes, and how to validate it with objective data,” said Wang Zhong. Clients are now more concerned about implementation, controllability, security, and integration with existing business processes.

Furthermore, the importance of knowledge engineering is significantly rising. Wang Zhong pointed out that as the gap in model layers continues to narrow, the key determinant of AI agent effectiveness has shifted to knowledge construction at the Agent layer. In the past, enterprises often equated knowledge engineering with simple document library construction and were unwilling to pay extra for it. However, “raising Lobsters” has made them realize that merely reading documents is far from enough—a large amount of experience must be infused, and this experience comes from knowledge engineering. Therefore, when constructing scenarios now, enterprises are generally more willing to invest resources in building general industry knowledge bases or enterprise-specific knowledge bases to prepare for the application of other scenarios in the future.

The <2026 iResearch · Central State-Owned Enterprises AI Agent Implementation Progress Report> also shows that the success rate of AI agent projects within central state-owned enterprises was about 70% in 2025, lower than the average for traditional IT projects. Among these, data and knowledge quality have become the primary reasons for failure. “Last year, many central state-owned enterprises built AI agent platforms based on open-source products, but open-source products were essentially blank in terms of knowledge governance, which became a fatal issue in implementation,” said Zhang Yang, co-founder and chief analyst of iResearch. He believes that knowledge governance should be a prerequisite and standard for AI agent implementation.

04 The Second Half Has Just Begun

AI agents are still in the process of value verification and are far from widespread adoption. A survey by Deloitte in August–September 2025 of over 3,200 global enterprise executives revealed that only 25% of enterprises have deployed AI agents in production environments, with implementation cycles stretching from an initially estimated 3 months to 18 months. However, rapid technological advancements in the first half of this year are bringing profound changes.

The industry is proposing increasingly complex dimensions for AI agent implementation, shifting from initially focusing on token consumption to looking at coverage rates, the number of core system-connected tokens, and the efficiency of agents autonomously completing closed loops within enterprises.

In terms of coverage, early adopters have shifted from isolated deployments to broader replication. Wang Zhong from Zhongshu Xinke stated, “Last year, clients mostly selected a few small scenarios for closed-loop pilots to achieve basic efficiency improvements. Now, most clients have entered the stage of comprehensive business chain review, assessing business scenarios holistically and formulating phased construction plans for this year, next year, and the year after, advancing in a coordinated manner.” He believes that due to financial, talent, and time constraints, the popularity (translate: prevalence) of AI agent applications among small and medium-sized enterprises in second- and third-tier cities remains low, and it will still take one or two years for the To B market to reduce costs.

Some high-value scenarios are emerging more rapidly. Xin Zhou from Baidu Cloud cited an example: a couple of years ago, AI agent technologies and products had not yet reached production-grade availability, with many being low-quality scheduling tasks, such as using large models for data cleaning or tasks that could be done with smaller models. However, with advancements in model and AI agent technologies, a batch of truly valuable applications has begun to emerge. For instance, Baidu Famou improved container handling efficiency at a port, outperforming the client’s original system.

In the maritime and ocean industry, Zhongshu Xinke is collaborating with Changhang Group and Wuhan University of Technology to develop an intelligent agent for voyage mission planning. When a ship sets sail from Chongqing to Shanghai, the intelligent agent will calculate the most economical speed and fuel plan in real-time based on hydrology, waterways, lock queue times, port service replenishment, loading and unloading conditions, etc. 'This used to be the captain's job, and the direct benefit is fuel savings,' said Wang Zhong. Currently, they have already penetrated two core business areas of shipping companies, interfacing not just with mid- and back-office departments but also with core operational departments like the Ship Management Department and the Technical Department. However, there is still a gap to cover the entire chain of core businesses. 'I think we should be able to go deeper in another year or two.'

In the industrial sector, product design and production have always been considered the most core links. Zhizhen gave an example: in the past, drawings had to be made manually, but now it's Text to CAD/CAE, allowing a 'pretty good-looking' transformer to be designed in just a few hours. However, the challenge lies in 'constrained random design'—industrial design must consider hard constraints such as production line capabilities and inventory to ensure that the product not only looks good but is also manufacturable. 'Product design is divided into two parts: the front-end drawing (CAD/CAE) and the back-end management (BOM generation). Currently, AI applications in the BOM part are fine, while the CAD/CAE part is still somewhat challenging, but it should be okay in the second half of the year.'

Zhizhen said they are also exploring integrated design and production. In the past, CAD design, production, and other links were isolated, with inconsistent underlying data models. In the future, it will be model-driven, with industrial large models uniformly defining product models, factory models, and industrial models, and business processes running based on human-machine interaction. However, this requires first cleaning up the existing data to a high quality, which involves significant investment.

Yu Youping from Zhongguancun Kejin also introduced that in automotive intelligent marketing scenarios, customers used to purchase 'efficiency tools,' but now intelligent agents can deliver business results such as 'conversion of high-intent leads.' In a project where a certain automaker used a large model outbound calling intelligent agent for customer outreach, they delivered result metrics such as marketing call connection rates, high-intent customer dispatch rates, and customer WeChat addition rates.

The iteration of industry implementation logic has also driven a comprehensive upgrade in the development strategies of intelligent agent vendors. Vendors have generally shifted from 'doing everything' to focusing on specific industries and scenarios. Zhongshu Xinke is concentrating its efforts on three core product lines: maritime and ocean, education, and bid review, aiming to deepen and excel in these areas, with plans to grow its business by about 50% this year. Zhonggong Hulian also emphasizes focus, with the goal this year of delving deeply into several very specific large clients and scenarios to create value that clients 'can't live without.'

Meanwhile, FDE (Field Deep Delivery) capability has become the core competitiveness for the large-scale implementation of intelligent agents. An Love Analysis report shows that 75% of central and state-owned enterprises place greater importance on suppliers' engineering capabilities, such as on-site research, business diagnosis, and closed-loop delivery.

To this end, major vendors are strengthening their FDE systems. Baidu Intelligent Cloud has built a professional FDE team, complemented by AI FDE intelligent assistance capabilities, to go deep into the front lines of customer businesses. Zhongguancun Kejin has launched an internal AI engineer certification system, reshaping capability models for key positions like FDE, and cultivating a high-end composite talent echelon (talent echelon) with deep industry Know-How and engineering implementation capabilities to ensure the large-scale and high-quality delivery of intelligent agent projects. Zhonggong Hulian also relies on the FDE model to solve implementation challenges in complex industrial scenarios. 'You simply can't do it without on-site presence now,' said Zhizhen.

As intelligent agents move from pilot trials to wider replication and accelerate from peripheral businesses to core production links, the second half of the intelligent agent story is just beginning.