In 2024, AI Agent has become the hottest term in the industry. Almost all AI companies are talking about Agent, all large models are integrating Agent, and all developers are trying to "make their own Agent".

The question is, can the Agent that everyone is talking about really be used?

Recently, ByteDance's "Button Space" was launched in internal testing, which gave me the first glimpse of the serious thinking of the domestic team on Agent. It is not a simple assembly of "putting a shell on a large model and assembling a tool chain", but a set of AI work systems that truly aim to "complete tasks" and are oriented to product delivery .

If you are interested in the idea that “AI is not just about chatting, but about doing work”, I suggest you read this article.

1. The dilemma of AI Agent: The model is stronger, but the experience is worse

ChatGPT has put the big model in the spotlight, but it has also created an industry illusion: as long as the model is strong enough, all problems can be solved.

But the reality is that the big model has been upgraded, but the user experience has not kept pace :

• Complex tasks are difficult for models to solve with just one question and answer

• It is difficult for users to know how to ask questions

• After multiple rounds of interaction, the model often "loses memory" or speaks nonsense

• The integration of tools is more complicated and not as fast as manual work.

Many "AI applications" are actually just a GPT shell with a few buttons and calling interfaces, which are far from the "autonomous thinking + task execution" capabilities that an agent should have.

OpenAI launched GPTs and Assistant API, trying to open up the Agent application path. There are also many imitators in China, but most of them remain at the tool level, and the product logic and scenario polishing are very rough.

ByteDance’s “Button Space” made me see something different: it tries to truly “use AI as a human” rather than “let people learn to use AI.”

2. What is "Button Space"? In a word, it is an AI workspace, not another AI chat room

You can think of it as a combination of Notion, Slack, and GPT, but the core is: it is oriented towards "tasks" rather than "conversations" .

In "Button Space", each Agent has a clear "responsibility". It is not an all-purpose question-answering machine, but an "intelligent employee". For example:

• Some Agents are responsible for competitive product analysis

• Some Agents do user research reports

• Some Agents do A-share data analysis

It's like a group of virtual employees you configure in your enterprise, each of whom has his or her own skill stack, tool chain, and workflow, and can get things done around a goal.

In specific tasks, the Agent does not directly give a result, but goes through a complete process of "understanding the goal - planning the path - calling the tool - outputting the result". This process is not pieced together by prompts, but a logically clear, debuggable, and iterative working system.

This is the real implementation method of Agent.

3. From “exploration mode” to “planning mode”, AI needs to think and plan like humans

Currently, "Button Space" has tested two types of Agent workflows:

Exploration mode: It does whatever you say, suitable for divergent tasks

Planning mode: Understand your goals first, confirm them clearly, and then execute them

The design of the planning model is very critical. This is actually a localized implementation of the idea of ​​OpenAI Deep Research: ask clearly about the goal before taking action, do not guess or rush, and improve the quality and stability of the output.

We tested a "competitive product analysis" task. In exploration mode, the agent can quickly list several competing products and compare their functions. However, the analysis dimension is coarse and key information is easily missed.

In planning mode, it will proactively ask about your competitor type, target users, and analysis depth before starting to build structured outputs, including:

Competitive product feature matrix (table)

Summary of user feedback (from various platforms)

Differentiated opportunity suggestions (based on user profiles)

More importantly, this process can be adjusted repeatedly, and you can modify the goal at any time during the process, and it can respond flexibly. This interactive experience is obviously more like "co-creation" than a question-and-answer AI chat.

4. The integration of data and tools is the watershed of agent capabilities

The core reason why many Agent projects are difficult to implement is not insufficient technology, but "data fragmentation" and "ineffective tools".

Agent is not a stage for displaying models, but a central control center for data and tools. ByteDance has done a thorough job in this regard:

Direct data connection : You can directly access Feishu documents, tables, and workflows

Tool registration standardization : integrating third-party tools through the MCP protocol

Multiple output results : support automatic generation of documents, charts, and visualization components

For example, if we test a task "Analyze CATL's expansion opportunities in North America in the next three years", a complete result includes:

GPT summarizes the market background (large paragraphs of language)

Automatically generated North American market sales trend chart (based on financial report data)

Timeline of corporate cooperation trends (graphic display)

Finally, package it into a report document and export it to Feishu with one click

This is not a scenario of "AI generating text", but a scenario of "AI completing work".

This scenario implementation capability is the lifeline of Agent products, and is also something that many "model packaging tools" on the market do not have.

5. “Expert Agent” is the real value entrance

General Agents can do more and more things like GPTs, and what really differentiates them is “expert Agents”.

The so-called expert does not mean that AI can imitate the speaking style of an expert, but rather "structure and encapsulate the expert's knowledge, experience, and process into an Agent model."

Kouzi Space currently has two relatively mature expert agents online:

User research experts : can conduct in-depth analysis of interview texts, refine user needs, and generate product recommendations

Huatai A-share Assistant : Based on Huatai Financial Database and public data, it can analyze corporate trends and predict influencing factors.

We conducted a case test: we asked user research experts to sort out the needs and make product recommendations based on the original interviews with 5 users. It automatically identified the users' emotions, motivations, and pain points, and output a suggestion table based on the product design dimensions, which is almost close to the work level of a mid-level PM.

The replicability of this expert agent is an important growth point for agent products in the future.

6. Finally, let’s talk about one point: Why did ByteDance do it?

The implementation of AI Agent is essentially a coordination issue among the four elements of “product, model, data, and tool”.

ByteDance has a unique advantage: data, tools, product sensitivity, and knowledge of “how to build a platform.” Over the years, the Button platform has accumulated more than 2 million AI application data, with enough real needs and failure lessons behind it.

In this context, "Button Space" is more like an upgrade after systematic reflection - from "AI applications blooming in a hundred flowers" to "building a working AI talent system."

This is not a demo in the laboratory, but a product-line-level product.

Last words

90% of the Agent products on the market today are still at the "building blocks" stage, combining large models, tool chains, and plug-ins in the hope of automating some tasks. However, there are not many that truly build complete workflows and controllable behavior systems based on user needs.

"Button Space" is not the end, it just shows one thing:

Agent is not a function, but a new interaction and collaboration paradigm.

If GPT is the next generation of search, then Agent is likely to be the next generation of "work operating system".

Whoever can truly follow this path will be able to reconstruct part of the knowledge workflow. ByteDance’s move is not perfect, but it has taken the right direction.