Following Manus, another general-purpose AI Agent product in China has come out (see Manus introduction:https://www.woshipm.com/ai/6191997.htmlThat is, Coze Space, launched by ByteDance on April 19. The platform is currently in the public closed beta stage, and an invitation code is required to activate it. After I got the invitation code, I was trying to play with it during the May Day holiday. So, how does the buckle space product perform, what are the applicable scenarios, and what are the advantages and disadvantages? In this issue, give me time for a song, and I will introduce you to “Button Space” in detail.
Recap: What is AI Agent?
In the article introducing Manus, we also briefly introduced the concept of AI Agent, which we will review here.
If you want to talk about AI Agent, this classic diagram is indispensable. For everyone’s convenience, I put the original picture and the Chinese translation I made together.
In simple terms, an AI Agent is an intelligent entity that can sense the environment, make autonomous decisions, and perform actions. Unlike traditional AI systems, it can not only answer questions, but also have memory capabilities, the ability to independently plan and disassemble tasks, and the ability to call various tools, so that it can actively complete a series of complex tasks. To put it simply, if a large language model (LLM) is compared to a “super brain”, then the AI Agent is equipped with “hands and feet” and “tools” to allow it to take the initiative like a human, rather than just answering questions.
How the buckle space completes the task
When I was thinking about a short video on the same topic, I was already thinking about how to better introduce the button space. Finally, I thought of a trick, which is to let the button space introduce itself. So, I sent the following prompt to the buttonspace:
I’m preparing a short video introducing “Buttonspace”, and I need you to provide me with the corresponding materials, and finally form a concise and clear report, and generate an interactive web page based on the report.
After receiving the task, the button space begins to be processed. The first is task disassembly, and Buckle Space believes that “first of all, we need to understand the basic information of ‘Button Space’, such as what type of space it is, what characteristics and functions it has, etc.” Once the materials are collected, begin to “organize this information into a concise and clear report and save it as a Markdown file.” An interactive web page is then generated based on the report. ”
Finally, the whole process of generating Markdown files and web pages, which took a long time, taking a full 10 minutes. Next, based on the introductory materials generated by the buckle space, I will put it directly on to you.
How to introduce clasp space
1. Basic information
Coze Space is a general-purpose AI Agent platform launched by ByteDance on April 19, 2025, with an official website address of https://space.coze.cn. The platform is currently in the public closed beta stage, and an invitation code is required to activate it. The word “space” in its name not only expresses that the platform provides a highly open and customizable collaboration environment, but also symbolizes that users can build their own intelligent workspace here, flexibly combining various AI agents to complete different tasks.
2. Functional characteristics
(1) Dual-mode collaboration
Exploration mode: AI responds autonomously and quickly, suitable for time-sensitive tasks, such as querying the latest information and obtaining real-time data. Users only need to enter task instructions, and the AI will automatically complete each step and quickly output results. For example, if a user enters “query the weather for the next 5 days, formulate a travel plan for a 5-day tour in Hangzhou, describe the specific travel route, and generate pictures of each attraction to give outfit recommendations”, the AI will independently plan and execute tasks, call the Ink Weather interface to query weather information, call the map to query planning information, and finally call the image generation interface to generate scene pictures.
Planning mode: AI deep thinking and execution, specializing in high-complexity projects. After the user puts forward the request, the AI first gives the task processing plan, and then starts to act after the user confirms it. For example, when formulating large-scale project plans and in-depth data analysis and other complex tasks, the planning model allows AI to think and analyze more deeply and provide more complete solutions.
(2) Strong task handling ability
Task automation: It can automatically analyze user needs, disassemble them into multiple subtasks, and independently call browsers, code editors and other tools to execute tasks, and finally output complete task reports, such as web pages, PPT, Feishu documents, etc. For example, creating a research report on an industry can automatically complete a complete set of tasks such as data collection, analysis, and organization.
Multi-agent collaboration: Multiple agents within the company can cooperate with each other to complete complex tasks together. For example, when making travel guides, different agents can be responsible for tasks such as querying attraction information, planning traffic routes, and finding nearby food, and then integrating and outputting the results.
(3) Rich plug-in integration
The first batch of more than 60 MCP modular capability plug-ins that integrate Feishu multi-dimensional tables, AutoNavi maps, image tools, speech synthesis, etc., will also support developers to release custom plug-ins through the “Buckle Development Platform” in the future, infinitely expanding the capability boundaries of Agents. These plug-ins cover information reading, travel, efficient office, shopping and consumption, finance, text processing, image and video, audio and music, knowledge management, data analysis and many other aspects, so that Buckle Space can meet the various needs of users in different scenarios.
(4) Expert-level agent ecosystem
Provide professional agents in a variety of fields, such as “Huatai A-share observation assistant” can generate daily stock market morning reports and answer stock analysis questions, “user research expert” can assist in in-depth analysis of user research data, and “e-commerce super brain” can analyze user behavior paths in the e-commerce industry, etc., providing convenient professional services for users in different fields to meet the diverse needs of different users.
3. Technical support
(1) Bean bun 1.5 model
Based on ByteDance’s Doubao 1.5 model, which contains 208 “sub-experts” and adopts a MoE hybrid expert architecture, these “sub-experts” are good at different tasks such as programming, vision, and logical reasoning, and can be dynamically combined to meet complex and diverse user needs. In tests such as the programming competition codeforces, its performance is benchmarked against GPT – 4.
(2) Volcanic engine
With the technical support of Volcano Engine, the platform’s powerful computing power and stability are ensured, and it can quickly process large amounts of data and complex tasks, providing users with a smoother experience.
(3) MCP protocol ecology
Similar to the HTTP protocol of the Internet, the MCP protocol allows third-party developers to access plug-ins, such as weather query, image generation, etc., to quickly expand the capabilities of agents and enrich the functions and application scenarios of the platform.
Fourth, advantages and disadvantages
(1) Advantages
Strong ease of use: AI can be driven to complete complex tasks through natural language interaction, without the need for users to have programming foundation or perform complex operations, and ordinary users can easily get started, greatly reducing the threshold for AI application development and use.
Efficient and convenient: It can automatically complete tedious tasks such as task disassembly and tool calling, and quickly output results, saving a lot of time and effort, and effectively improving work efficiency. For example, when conducting competitive product analysis, Buckle Space can complete the analysis report within 11 minutes and automatically deploy it online.
High scalability: The rich MCP extension plugins and future support for custom plugins enable the platform’s functions to be continuously expanded and enriched, better meeting the individual needs of different users.
Professional Service Support: The expert Agent ecosystem provides professional services and solutions for users in various fields, enhancing the practicality and credibility of the platform.
(2) Insufficient
Limited capabilities of the pedestal model: Currently, relying on Byte’s own bean bag model, it is sometimes inferior to top players such as GPT and Claude in terms of depth, logic and creativity of plain text content, and the output content may not be perfect.
Stability issues: The AI Agent execution chain is long, there are many intermediate links, and the probability of errors increases accordingly, and sometimes task execution fails.
Hallucination Issues: May produce AI hallucinations of “serious nonsense”, resulting in inaccurate or reliable output results.
The ecosystem needs to be improved: Although the official MCP expansion is provided, it is far from rich and powerful compared to the needs of the real world, and the plugin ecosystem needs to be further enriched and improved.
5. Application scenarios
(1) Office scene
It can help users complete various tasks such as writing documents, making PPTs, data analysis, meeting arrangements, and email processing, and improving office efficiency. For example, users can ask Buckle Space to write research reports, compare the advantages and disadvantages of different platforms, output Feishu documents, generate forms, etc.
(2) Learning scenarios
Provide students and educators with auxiliary teaching and learning services such as course material collection, sorting, courseware production, and study plan formulation to help the development of education.
(3) Life scenes
For example, making travel plans, querying information, booking hotels and air tickets, health management and fitness guidance, etc., to facilitate users’ daily life. For example, users can ask Button Space to generate travel guides, plan itineraries, etc.
(4) E-commerce scenarios
It can conduct product recommendations, user behavior analysis, market research, etc., to help e-commerce companies improve sales performance and user experience. E-commerce platforms can import product information and user portraits into the button space, and then combine user purchase records and natural language demands to accurately recommend products that users may be interested in.
(5) Financial scenarios
Such as stock analysis, financial data interpretation, risk assessment, etc., to provide decision-making support for financial practitioners and investors. “Huatai A-share observation assistant” can generate daily stock market morning reports and answer stock analysis questions for users.
6. Comparison with similar products
(1) Comparison with Manus
Traditional chatbots are more like “question and answer machines” or “information inquirers”, mainly relying on their own pre-stored knowledge to answer questions, while Buckle Space, as an AI agent, has the ability to call external tools and independent planning and execution, and can really “check the map” and “look at the weather”, etc., connect the real world, and use various tools to solve practical problems.
Finally, talk about personal feelings
Finally, I would like to make a few comments. Let’s talk about the surprise part, the most significant thing is that compared with the traditional AI Q&A mode, the buttonspace is directly from the initial task prompt to the finished Markdown report and the generation of interactive web pages in one go. That is, AI almost completely takes over the process of task disassembly, execution, and tool calling, which is also the biggest breakthrough in the AI Agent model.
Then there are the shortcomings. First of all, it is time-consuming, and it took 10 minutes and 28 seconds to complete the task above me, compared to Deepseek’s 2 or 30 seconds of thinking process. Then there is the stability aspect, the above finished product was actually successful until I retried it for the 4th time, and the first 3 times failed when generating web pages, which shows that for complex tasks, the stability of the button space needs to be improved (this point is also mentioned in the introduction of the button space itself). The last and most important thing is that for the details of the finished product, the current support of the button space is not good, such as the above web page, if you want to modify the detailed performance, you can only start anew.
But in general, as a product that has only been released for less than a month, these shortcomings are engineering problems and can be continuously improved with the iteration of the product. I think that any new thing is neither blindly touted nor blindly denied. All products will have shortcomings in the initial stage, but more importantly, we need to see which ones represent the future development direction and which are the areas to be improved in the development process. And Buckle Space, or AI Agent, represents a more thorough human-AI collaboration model. It is still worth looking forward to what kind of results it will develop in the future.
