Google Deep Research VS Manus: Guan Gong vs. Zhao Si

There was no official email reminder from Manus, but I could use it just by logging in yesterday.

A simple test compared the research and analysis capabilities of Google Deep Research with 2.5 Pro on the same topic.

Given a very simple task "API access to Telegram channel content"

Manus's operation is very exaggerated. It directly starts a Ubuntu virtual machine and then completes all operations in the sandbox, including controlling the browser and calling command line tools.

I can only say that this wave is amazing and very imaginative.

Then comes the acquisition of content, which completely simulates the information of manual web browsing.

The files are modified and edited directly in the middle. But no thought process is given for the task.

After experiencing it, I feel that there are still many problems.

The first is the compliance of the content. The official website of manus is unlikely to be open for domestic use. This kind of proxy Internet access model is too borderline.

I also heard recently that they are seeking cooperation with Alibaba.

But if the interactive execution logic is the one demonstrated above, I feel that it doesn’t matter much who you join with or who you join with.

Because major domestic platforms are now isolated islands and are not connected to each other. They either require scanning codes or SMS verification, so it is basically impossible to obtain valuable content.

Then there is the virtual machine model, which is still a layer of separation. It may be OK to simply process commands, but it is still a long way from productivity. After all, the real user-oriented tools are still physical application software such as Excel, PS, and IDE, not to mention the processing of document resource data. It is impossible to upload and download through a browser, right?

Manus summed up his overall feeling in four words: "simple and crude". As for whether it is effective or not, the conclusions may be different for different tasks. For this task, the degree of completion is actually acceptable.

However, compared with Google Deep Research with 2.5 Pro, it is really a dwarf compared to the giant, and is not in the same league at all.

While Manus was starting up the virtual machine and manually clicking on the browser one by one, Google Deep Research was directly connected to its own database. Combined with the super context of 2.5 Pro, it produced a 5,000-word report in two minutes.

Almost a small paper has come out.

Today's agents are still fake agents. They are definitely not agents that can control a browser, crawl some data, and connect to a model.

A truly universal agent must have rich data support, be built on a powerful visual model, connect various productivity tools, and implement an AI operation layer on physical computer and mobile phone operating systems. As of now, it seems that only Google can complete this universal agent.