Don’t let “independence” become a shackle for intelligence

Today we continue to talk about intelligent agents along the lines of Manus.

In the past few days, the public opinion about Manus has experienced many reversals. There is even a team of 5 people who spent 3 hours to fork an open source version of Manus  - OpenManus.

On March 11, Manus announced its collaboration with Tongyi on its Chinese website.

Screenshot of Manus official website

On the same day, OpenAI launched several agent APIs and an agent development SDK at its latest product launch conference. Unlike Manus's route of directly creating general agents, OpenAI provides developers with tools to create agents. The button mentioned earlier is somewhere in between.

Kevin Weil (OpenAI's CPO) defined the agent at the beginning of the press conference as "A system that can act  independently  to do tasks on your behalf". At the end of the video, he even asserted: "2025 is going to be the year of Agent".

If training LLM is likened to training students through theoretical knowledge in college, then intelligent agent technology is like allowing these college students to enter society and start working.

Regardless of whether Manus is universal , let’s first see whether it is truly independent .

01

Travel Planning

There is a travel planning case on Manus’ official website, and it just so happened that I was recently planning a family trip.

I used GPT to make travel plans more than a year ago, and my LLM at that time did not have reasoning ability.

This time, I tried some travel AI applications, such as Layla, but the results were not ideal.

Afterwards, I used eight major AI tools from China and the United States to assist in planning: GPT, Claude, Perplexity, Grok, Secret Tower Search, Tencent Yuanbao, Tongyi, and Get Notes. Each of them is an expert in both searching for information online and reasoning.

Manus uses Claude and Tongyi Qianwen's LLM at the bottom layer.

The playback video of the Manus case is nearly 20 minutes long. In summary, it first draws up a To-Do list with seven major sections, then conducts online searches for different aspects (attractions, Japanese, travel tips, etc.), and finally integrates the information to draw up a complete itinerary and makes a travel guide in HTML format.

The whole process of thinking seems to be meticulous, the research is detailed, and the final travel guide is also quite amazing. But judging by my personal standards, this travel plan is almost completely unrealistic .

I think the problem lies in two words - " information ".

Making a travel plan is basically a comprehensive planning task. When we first form the idea of ​​traveling, there are countless possibilities (or uncertainties), and the core role of information is to reduce uncertainty. Making a travel plan is a process of constantly collecting and processing information, and gradually narrowing the decision space.

The amount of information Manus processes when making travel plans is far from enough .

Lack of information

The information that Manus lacks can be roughly divided into two categories  :  user demand information and evaluation of various elements in the itinerary . Both are subjective information .

In the case of Manus, the user requirements are as follows.

Perhaps only an ESFP would be so brief in providing information about their travel plans, not even mentioning hotel preferences and dining requirements. As expected, the itinerary Manus created did not include specific hotel and restaurant recommendations.

I counted 11 searches during the planning process, each returning more than 20 pages, but it seemed to read only one of them each time. Even assuming it read all of them, that's only a little over 200 pages in total.

When I used Grok's deep search function to make a plan, it read 199 web pages, which is similar in magnitude. Secret Tower Search has the largest number of reference web pages, reaching 515. The number of reference web pages of the other AI tools is only about 10.

200 web pages may seem like a lot, but for travel planning in an unfamiliar city abroad, this information may be enough to collect objective information, but it is far from enough for evaluative subjective information.

For example, the hotel address is an objective fact, while the quality of the hotel is a subjective evaluation; the restaurant opening hours are objective information, while the taste of the food is a subjective experience; whether there are cherry blossoms in a park is an objective fact, while whether it is the best viewing spot is a subjective judgment.

Personally, I think that subjective information has a much greater impact on a travel plan than objective information. The lack of both types of subjective information makes it impossible for AI to develop a travel plan that truly meets the requirements and is detailed enough.

Capabilities required by intelligent agents

I think intelligent agents need to have some abilities to deal with the lack of subjective information.

Ask proactively

The initial requirements I provided to AI were also very simple, but these were not all my requirements, they were just what I thought of at the time. It takes a process to perfect the requirements.

While it appears from other demo videos that Manus allows users to add information as it works, this relies on the user actively providing it.

After the emergence of LLM, many people believe that the ability to ask questions is the key to using LLM well. I think in the field of intelligent agents, the ability of LLM to ask questions is also the key to determining the quality of intelligent agents. Intelligent agents need to be "smart " enough to actively ask good questions and inspire users to provide relevant information.

Breaking down information silos

In the case of Manus, it encountered human-machine authentication problems when trying to access TripAdvisor and turned to other websites. Although TripAdvisor’s human-machine authentication is not difficult to solve, it reflects that AI may not be able to obtain some useful information for various reasons.

Among the eight AI tools I use, there is a distinction between China and the United States. Although I input Chinese commands, the American AI tools search for English information.

Even though they are both Chinese AI tools, the information they reference is not exactly the same. All 14 pieces of information referenced by Tencent Yuanbao come from WeChat public accounts, while 5 out of 7 pieces of information referenced by Alibaba Tongyi come from Quark (also an Alibaba product).

Today, the most trusted source of travel guide information in China is probably Xiaohongshu. Even if I explicitly asked Yuanbao or Tongyi to refer only to Xiaohongshu, they could not do it.

This is why I use Get Notes. Get Notes can transcribe the content of Xiaohongshu notes through links and store them in the knowledge base, so that Get Notes can refer to this knowledge base and use the embedded DeepSeek R1 to make travel plans.

Information Synchronization

It is highly likely that travel plans will not be finalized at once and may still change until the trip actually takes place.

After the agent completes a version of the travel plan, we will still actively or passively receive travel-related information, which may affect our judgment, needs, expectations, etc.

For example, Xiaohongshu may push a food note about a travel destination, which makes us want to try a certain restaurant. We would like to incorporate a meal at this restaurant into our travel plan. The agent needs to be able to synchronize information and revise the itinerary in a timely manner.

In addition, family trips often involve multiple people, and the agent needs to be able to synchronize information among all of them without requiring one person to act as a porter of information.

Conclusion

Let’s go back to the question at the beginning of the article  - whether Manus is independent .

If we follow Manus's case study, it is indeed very independent. After entering the task instructions, it can complete a series of operations and give the final result without human intervention. Perhaps this is also the fantasy of many people about intelligent agents, that anything can be generated with one click.

But for relatively complex tasks, perhaps the agent should not be completely independent. The criterion for measuring an agent may not be how many different types of tasks it can complete at the touch of a button, but whether it knows how to cooperate with (different) people to complete the task.

Intelligent agents need to be sensible, and more importantly, they need to understand you .