I'm sure no one will forget the "Tragon fever" at the beginning of the year!
OpenClaw stood out, with a 72-hour GitHub star rating of 680,000, and "Have You Raised a Dragon?" was once a buzzword among developers. It was followed closely by more than 30 manufacturers.
And then what?
On April 4, Anthropic formally cut off OpenClaw's call authorization for Claude's subscription limit, and the heat hit the first wave of cold water. Shortly thereafter, more than 135,000 cases of OpenClaw were found to be exposed to a lack of protection on the public internet, and frequent supply chain attacks; the National Intellectual Property Office issued a security alert; and in July, bytes bouncing bean bags and Ariot were required to shut down personalized AI Agent functionality.
AICPB data shows that in April, OpenClaw's traffic declined by 50.67%. In the community, "crowding out tutorials" are more common than tutorials. Even the most recent update, OpenClaw 2.0, has lost much of its attention.
The tide is back and all that is left is real gold.
China Connected is one of them. It is based on the self-developed Agent Harness platform and rebuilt from its underlying architecture a set of workplace-oriented AI-smart bodies.
UniClaw has evolved from UniClaw, which features sandbox isolation, data offline, open-boxing, and full-scene tasks. These are the skills that were neglected during the heatwave and became the most scarce after the downturn.
It is not easy to become a true "digital employee".
Are all kinds of smart and well-behaved employees really alive, or is it just another chat room with Agent's outfit?
We don't look at advertising to see how effective it is. We put it in some of the most common types of office tasks and see how this "digital employee" level is.
Experiments: What can digital workers do?
When I opened the digital staff, I found that the main interface was very easy to use, simple, and easy to understand for users familiar with AI tools.
The coverage capacity is also very comprehensive, with two models of office life and code development covering the five business areas: "Office, R&D, life entertainment, education, learning and financial investment".
In each area, there are a number of default prompter templates for common tasks such as documentation, data analysis, research, game development, performance optimization, etc. that make it easy for each user to build a prompter.
First impressions are good, then try!
The version I started testing supports the DeepSeek V4 Flash and GLM 5.2 models, and the test cases were done using the DeepSeek model.
Test One: In-depth study of a sentence
When it comes to the workplace, publishing requests is often a single word, followed by a whole series of workflows.
So in the first round, I decided to increase the difficulty to see if this number of employees could start from a single need and run a whole process of research on their own.
Testing tips: "Research the financing trends and key players in the AI Agent market in 2026 and produce a research report that includes data graphs, focusing on differences between China and the United States".
This test is cross-capacity module synchronization: extracting raw information from web searches, extracting structured data after cleaning, generating visualized charts with data analytics capabilities, and finally producing a format specification, logically complete interactive reporting. The entire process involves network search, data analysis, depth study, and file generation of four modules, with any loop in between affecting the final delivery quality.
In the process of searching the data, it first performs a large number of extensive searches, and in the process cross-validates the authenticity of the information obtained at any time, which gives me relative confidence in the correctness of the results.
In the process of constructing visualization charts, the system demonstrated its multi-modal recognition capabilities. By scanning and confirming whether the results were in line with expectations and making targeted code adjustments based on the image recognition results.
Finally, a comprehensive study report in HTML format was produced, divided into six main sections, with detailed content, graphics that were easy to read and emphasized.
This task may seem simple, but it is not.
It's an open-source task, and my hints don't tell me exactly how to do it. It requires a complete process to be planned by itself, and to choose the right tools at every step. This tests the core capabilities of "autonomous planning and execution" in the Harness architecture.
In addition, the task also tested the coding and multi-mode recognition capabilities of the intelligent body.
Test Two: Full-time Assistant on call
I've chosen the WeChat Clawdbot binding test.
Test Tip: "Please open your browser automatically to access content related to searching, crawling, and AI or technology, generate a diary and send it to me. Now do it once to make sure the task goes smoothly".
Browser access is one of the easiest scenarios for Agent to "roll over". Changes in page structure, verification code interception, asynchronous loading can disrupt the automation process.
I observed that digital workers encountered multiple blast windows and counter-crawling intercepts, but each time they tried a different way of circumventing it, never giving up directly.
Before constructing a script, it will actively verify the stability of the script, ensuring that there will be no failures in the future due to problems such as pop-ups. What's more, it will actively filter out the content it gets and avoid pushing dirty data.
In the process of constructing the script, an interesting detail emerged: the system found that the keywords in the original script matched the results of the "AL 0:3 BLG" competition, mistakenly judging it to be AI-related, and then located the problem and fixed it itself.
Eventually, I received a screened and integrated tech diary from WeChat. It can be executed multiple times, which is a significant reduction in the editorial department's selection work.
There is also an intuitive sense of security. When choosing a work path, the system will explicitly require user authorization, without violating the rights.
In code development, you can also choose to grant permission to digital employees, which includes three security authorization strategies: cautious, balancing, and letting go.
It's safe to go across the border.
To truly verify the availability of an Agent product, it is not enough to look at the successes in the demonstration video, but at its stability and fault tolerance under border conditions, abnormalities, and long-term operation.
This employee is a bit different.
Following previous observations, I have found that the number of people working in digital connectivity is indeed somewhat different from that of other AI office products.
If you want to deeply internalize AI capabilities into the productivity workflow, there are actually three questions that people are most concerned about: Is it easy to use, is it inadequate, and is data security unsafe?
The level of resolution of these three key issues, the basic level of satisfaction, and the process of solving them, also makes people mindful, reassured, reassured, let users know and know why!
Keep in mind, all channels are reached, missions are uninterrupted, progress is responsive
A qualified digital employee is able to achieve life expectancy, connectivity, and execution at any time.
It enables users to access through full channels and supports almost all the touch channels you can think of, including PCs, WeChat, Nails, Cloud Computers, 5G messages, voice, video integration, and other AI native communication inputs, making the capabilities of the smart body both convenient and affordable.
Regardless of the time and place, cross-sectional digital employees are able to achieve multi-channel wirelessness, tasks are never interrupted, and progress can be reported at any time.
Rest assured: Huge skills automatically awaken, work lives end efficiently
The biggest feeling I had was "I can't think, I can't do it".
Former Digital employees have accumulated more than 1,000 skills in the five business areas: "Office, R&D, life entertainment, education, learning and financial investments".
With the unlimited expansion of ecological skills such as business management, food and clothing, users can automatically dismantle tasks, match skills, and perfectly execute tasks in a single phrase.
Behind this is the MaaS platform, which allows each token to be produced and supplied, updated, stronger, more complete, so that each token has a value.
Supported by the Maas platform, mainstream open source models, including DeepSeek, Millennial Queries, GLM, and others, are able to make Day0 synchronous online. Top 5 models in the industry are fully covered and can provide intelligent routing.
The MaaS platform is capable of reducing the latency of words by more than 50% on average, increasing throughput by more than 2 times, and improving the maximum reasoning speed by 5 times.
On the capability certification level, I looked at MaaS, which has received 5 4+ and advanced certifications, such as Credit Union Model Services, Smart Bodies, RAG.
On top of this, it also built "Smart Body 114" skill-grading precision matching technology, breaking the bottleneck of skill numbers.
Security: Multidimensional security, reliable operation of central enterprises
In terms of security, you can always rely on the power of the Central Bank.
For instance, digital employees protect each link of input, reasoning, and output of model calls, intercept common attack risks, and support pre-evaluation, interception, and back-up process control.
Most importantly, VIS digital employees have a self-developed Harness architecture that enables autonomous control of underlying technologies.
Harness's core design is to integrate long-term memory with a self-reflective mechanism to make it more efficient for digital employees to perform the same tasks over and over again.
The enterprise version can provide a privatised cloud-side deployment program to achieve unified management of digital employee skills and ensure the security of enterprise production and operations.
With this set of central enterprise services operating system, the entire escort can be used by both enterprise and individual users.
Industry Skills, enabling digital workers to go from generic to specialized
As you can see, digital tools have good basic capabilities. But some professional scenarios require AI assistants not only to be smart, but also to really understand the rules and knowledge of the industry.
In this regard, the cross-sectional digital employees gain more reliable professional capabilities through the overlay of industry skills, prioritizing expertise, industry specifications and process logic for specific areas of work, so that the smart body reaches the level of "being good at using, coming out on compliance" in vertical scenarios. In one sentence, the system can automatically schedule the corresponding industry skills to complete tasks.
Industry Skills are currently in place, covering government, industry, and tourism, with each case in operation.
Public administration codes are confidential, low error tolerance, letters and regulatory number errors can be reversed, and the cost of artificial verification is difficult to eliminate. For this purpose, the Secretariat Intelligence has created six professional Skills, covering 15 categories of statutory public text judgments, writing scenarios, one-click audits, text and comparison audits, form audits, and intelligent evaluation through the entire process, which can be output with revised batches of Word manuscripts.
In a multi-dimensional presentation on "Policy Simulation" conducted jointly by China Link and Tsinghua University, the test found that its form-audit accuracy was 100%, content audit accuracy was more than 90%, and the effectiveness of auxiliary decision-making was reduced from a few days to 10 minutes.
Ship design has long faced difficulties in finding drawings, time-consuming layout reviews, and three pain points for repeat layout efficiency. The Ship Design and Intelligent Industrial Design platform created AI+EID Assisted Design Skill, using a "three-step" rebuilding process: automatically matching mother ship drawings by ship type, position, number of boxes, then identifying layout drawings and reviewing model names, line specifications in AutoCAD, and finally generating EID electrical principles according to the device line.
In practice, the time taken to generate map recognition and principles has been reduced to 5 minutes and 10 minutes respectively, and the system's map efficiency has been improved by about 40%, and the 2026 World Conference on Artificial Intelligence "AI Empowerment" has selected a case for marine innovation practices.
The information of the tourism scene is highly fragmented, visitors need to switch on multiple platforms, third-party approaches are poorly aligned, there are no official backbooks.
The program uses NFC scripting toys as a carrier, easy-to-touch second-level awakening, zero-download registration; connects executive units, scenic areas and compliance merchants to provide authoritative recommendations, integrates guideline explanations, smart questions, route planning, dining options, itinerary planning, and translates from "people looking for services" to "people looking for services" and can be replicated in museums, antique streets, non-monuments and other scenarios.
The landing paths vary across the three industries, but this is a landing paradigm of the "General Smart Body Framework + Professional Skills": general capability as a foundation, industry knowledge as a depth, and cross-sectional digital workers working as a unified gateway linking the entire capability spectrum from the general office to the professional scene.
I've been trying to find out.
The question I was most curious about before the test was: where did the excitement of working with an intelligent body come from?
While it has not yet "changed the world with one word", it has accumulated a steady and reliable degree of trust. For the past few years, AI has been able to run this long process of searching, analyzing, mapping, and reporting on its own; timely tasks to send selected daily newspapers to WeChat in a timely manner each day, encountering bulletproof windows in the middle, making mistakes and correcting them on its own; and in exchange for writing documents, drawing electricity, and making documentary guides, these professionals are still alive and well.
In general, the longer you use it, the more you can realize that it is already like an assistant who doesn't need to hold on and doesn't want to drop the chains.
Of course, AI tools are not enough to make a visual evaluation, and only real tasks can be used to find out.
The official website of the former SIG Digital employee:
If you're too busy with a bunch of repetitive, fragmented work, try it yourself and let the digital worker help you get through the day, and you'll have the answers for yourself!