What will the future form of universal agents look like?

After Manus became popular, people began to pay attention to a new concept:

General agent.

Some people think that this thing will completely change the direction of future technological development and subvert the way large-scale model products and humans interact; others fantasize that in the future, they only need to say a prompt word and AI can handle everything without having to do it themselves.

However, I think that general intelligent agent is not a new concept, and Manus is not the world's first AI general intelligent agent.

They simply define their product as a general artificial intelligence agent that connects thinking and action, with the characteristic of being able to handle various complex tasks. It can not only think but also give results directly.

So, don't be fooled by the fragmented information. If you do a little research, you will find that as early as 2024, there was already a lot of information about Universal Agent on the Internet.

What exactly is a General Artificial Intelligence Agent? Let me give you an example:

You are a student in the marketing department, and your boss suddenly asks you to submit a promotion analysis report of a new product on Xiaohongshu before 10 am tomorrow. In the past, you had to log in to the backend to import data, organize it into charts, put it into PPT, check the format, and finally hand it to your boss. The whole process is cumbersome and time-consuming.

If there is a general agent, just download the data, give it to it, and then say: Help me organize it into a report, visualize the data comparison part, and hand it over to the boss before 10 o'clock tomorrow morning.

Next, the general agent will automatically analyze the data, generate charts, summarize the results and make recommendations, and finally deliver the report to you on time; ideally, it can even help you make a to-do list and tell you what to do, and it will complete the rest of the tedious tasks.

Universal agents free people from repetitive tasks and allow them to focus on more important things.

I asked the AI ​​to look up its full definition for me. The AI ​​said:

A very smart AI, unlike a "specialist" who can only do one thing. It can understand complex needs, plan steps, and call tools. It is a "jack of all trades" that can automatically complete multiple tasks, and can respond to flexible scenarios just like a real person, truly achieving the leap from a single function to comprehensive assistance.

There are many different forms of general intelligent agents. The first one that comes to mind is the DingTalk AI assistant.

Friends who have used DingTalk should know that you can say to the AI ​​assistant: Check the chat history between me and Zhang San about plan A, and extract the key parts into text to form a to-do list.

It automatically searches for conversations, extracts key content, and organizes it into clear summaries. It can also add tasks that need follow-up directly to the to-do list, without having to look through them yourself.

I think this kind of tool has great potential. As a common tool in the enterprise, almost everyone has to use it at work, and in the process of using it, a large amount of data will naturally be generated and accumulated.

In this way, the role of AI assistant is obvious:

It can organize scattered data into useful information, and then directly connect them through various tools to turn them into something you can use directly. This will not only improve personal work efficiency, but also make the overall collaboration of the company smoother.

But it also has certain disadvantages.

Compared with products like Manus, Manus is more like a "connector of intelligent agents". It can connect agents with different expertise and coordinate these capabilities through a core agent.

DingTalk AI Assistant cannot do this at present. It is more about completing specific tasks independently, such as searching chat records, arranging schedules, etc. Some of these functions have not yet formed a deep linkage.

In other words, the DingTalk AI assistant is currently more like an efficient "single soldier", while Manus is like a "team leader" who commands multiple experts to work together.

Therefore, from the perspective of the depth and relevance of application scenarios, Manus has deep horizontal capabilities, especially in multi-task collaboration or complex process processing, while DingTalk is currently focusing on improving basic efficiency within the enterprise. If it can make further breakthroughs in task connection and intelligent collaboration in the future, it may be more competitive.

The second product form comes from the future evolution of large-model dialogue products such as Deep Research, Doubao, Kimi, Tongyi Quark and GROK3.

To better understand this, let me introduce Deep Research.

Deep Research is an enhanced capability built into ChatGPT, and is also a tool designed specifically for automating complex online multi-step tasks. It is highly similar to DeepSeek not only in terms of language style, but also in terms of functionality.

It can quickly find useful information from massive amounts of data, and can help you analyze, summarize, and even generate reports; you can think of it as a particularly smart research assistant that can handle most of the tedious work as long as you make a request.

For example:

You are an entrepreneur and want to know the trend of the health food market in the next three years. Just tell it this requirement, and it will search for information, sort out the key points, such as market size, what consumers like, how technology develops, etc., and finally give you a report.

But this is not the most critical thing.

Most importantly, when making a report, it can not only write text, but also make tables, codes, and even mind maps, so that you can understand the information more clearly.

However, its visualization capabilities are more like version 1.0. The functions are there, but the experience is not "amazing" enough. Why is it version 1.0? There are three reasons:

First, the basic functions have been implemented, but the details are not perfect enough. Most of the reports I have seen so far are in Markdown format , with text and simple structured data as the main content. In other words, the content presentation is relatively "plain" and lacks a more intuitive visual design.

Second, it is not very good at adding multimedia elements such as pictures and icons ; third, its interactivity is not strong enough.

A few days ago, I asked Kimi to help me generate a table. It was produced very quickly, but when I wanted to modify the content, I couldn’t do it directly. In the end, I could only give the prompt words again and again and generate it again and again.

This does not mean that their potential is limited. Version 1.0 is just a start. In the future, when we use tools such as Kimi, Doubao, and DeepSeek, their multimodal capabilities will become stronger and more interactive.

You can say to it: "Write a paragraph for me, and then make it into a poster image." It will immediately generate a designed image, and you can save it and use it as a poster directly.

Therefore, the development of these products will probably go through four stages:

In the first stage, it is a simple search and conversation tool that mainly helps you find information and answer questions. In the second stage, it becomes a "toolbox" that can write code and make tables.

The third stage is the all-round player, who combines text, pictures and chart capabilities to make the experience richer.

The fourth stage is a general agent with a high degree of autonomy and adaptability. It can automatically adjust the task process according to needs and even guess what you need next, truly becoming a general intelligent assistant.

This is what big models do. How they advance depends on how each company defines their core value.

Next, let’s talk about the third view, which is inspired by Flowith.

Flowith is a canvas-style creation platform that is intuitive to use, just like writing on a whiteboard.

You can create many nodes on it, each node represents a problem or topic, and you can connect them to form a thinking network. This design allows us to handle multiple tasks at the same time.

Flowith 2.0 integrates AI creation, knowledge management and services. It also has many practical functions, such as generating mind maps, calling Midjourney, and supporting AI models such as GPT-4 and Claude-3.

Flowith also has an "agent market" where you can share your own AI workflows and others can use them, so that everyone can learn and improve from each other.

What do I want to express after saying so much?

This type of product will start with a very simple entry point, and then gradually develop into a complete task flow or workflow based on the founder's understanding of the AI ​​product.

In this workflow, we can do whatever we want. In the end, the dialog box will become a universal intelligent assistant. In other words, as long as you ask a question, it can help you get through all the links and complete the task.

The fourth view comes from: Taskade AI Agents.

In layman's terms, Taskade AI Agents is a team of intelligent assistants in project management. They are driven by AI and automatically complete some repetitive or tedious tasks, saving you time and effort in team collaboration and project management.

You are planning a company annual meeting. You have to write promotional copy, sort out budget data, assign tasks to colleagues, and monitor progress. Doing it all by yourself is overwhelming.

Taskade AI Agents are like hired virtual assistants. When you ask them to write an annual meeting invitation copy, it will immediately generate one for you to modify and use. If you throw a messy budget spreadsheet to it, it will organize it neatly.

Tell it to "assign the venue arrangement task to Xiao Li", and it will automatically assign it and remind you of the deadline. If you ask "what else is needed for the annual meeting", it can also give you suggestions based on the situation. It's like chatting with a smart secretary, and you can also "train" it to make it understand you better and work more attentively.

So, the moment you create a task on this platform, Taskade AI Agents will immediately help you come up with some ideas, turn it into a task, and break it down into specific steps.

This is a bit like Notion, but different. The difference is:

Notion is like a flexible notebook, where you have to build the structure and input content yourself; whereas Taskade AI Agents are "alive" and will proactively help you generate content, divide tasks, and predict the next step.

Taskade is more oriented towards team collaboration and project management, with an AI assistant directly embedded in it; while Notion is more of an all-round tool, suitable for personal notes and databases, but the automation is not that strong.

In any case, it is like a general agent that accompanies you from the moment you create a task until the end of the task, helping you arrange every step properly.

The fifth general agent product form is: Zapier AI Agents.

What does it do?

Just say: Help me organize the new customer list and send follow-up emails, and it will immediately pick out the list from your form, write the email and send it out.

This experience is very smooth, mainly reflected in its chat box, which is similar to the feeling of chatting with Kimi or Doubao, but it has a special place:

When I say something, it not only replies, but also turns it into action directly; it can be used in more than 7,000 applications, automatically handling various tedious business tasks, and I can get it done just by moving my mouth.

Compared with Taskade AI Agents, the difference is obvious.

Taskade's AI assistant seems to be focused on working on its own project management platform, while Zapier AI Agents is a cross-platform "universal assistant" that is good at stringing together and automating work scattered across different applications.

Simply put, from the moment the order is given, it acts like a tireless assistant, managing the task from start to finish.

Its customer base is also quite broad. Officials mentioned that more than 50,000 companies have changed the way they work through AI functions. In terms of application scenarios, it covers various types such as entrepreneurs, medium and large teams, marketing and e-commerce practitioners, etc.

The last new form I mentioned is what I saw in a Tencent report in February this year. The name of the report is "DataLab: A Unified Platform for LLM-Powered Business Intelligence".

What is it mainly about?

DataLab is an intelligent platform that uses artificial intelligence technology to integrate various complex data processing and analysis tasks in the enterprise. Data experts can tell the platform their needs through simple conversations, and then the required content and charts will be automatically generated.

Simply put, DataLab is a very smart tool that makes data processing simple and efficient.

After reading the report, I understand that DataLab aims to build a large LLM (Large Language Model) platform and integrate multiple agents. Each agent has its own expertise, such as writing SQL statements, generating charts, and analyzing data.

Ultimately, you only need to ask a chatbot a question, such as: "Show me a bar chart of the products with the highest sales this year." The platform will immediately mobilize resources and generate the required charts for you.

Overall, I think Agent is about to enter a new stage of development.

Some agencies operate in a top-down manner, starting with projects and tasks, using a chatbot to connect multiple agents and allow users to use their products on one platform.

Either the agent is turned into an AI assistant, to which users can assign various tasks and let it manage and execute them; or a bottom-up , starting with simple searches and recording of small habits, and gradually building a complete automated system.

To put it bluntly, one large model dialog box was not enough before. Now you need to use one dialog box to mobilize multiple tools and solve complex tasks.