The Era of Universal Agents, Manus Takes a Step Ahead

Author｜Li Yuan

Editor｜Jing Yu

Manus, which has moved to Singapore, has not stopped its contemplation on universal AI Agents.

At the Stripe Tour held today in Singapore, Manus co-founder and chief scientist Peak (Ji Yichao) had a conversation with Paul Harapin, Chief Revenue Officer for Stripe in the Asia-Pacific region and Japan.

During the discussion, Manus AI revealed its recent operational data, stating that the Revenue Run Rate (RRR) for Manus AI has reached $90 million and is soon expected to exceed $100 million.

Xiao Hong from Manus AI specifically mentioned that the Revenue Run Rate refers to the current month's revenue multiplied by 12 and does not equal cash income. Many AI products offer annual payment options, which can only be counted as prepayments and cannot be recorded as revenue. "If we disclosed it in this [incorrect way], we could calculate a number greater than $120 million," Xiao Hong stated.

In addition to operational data, Ji Yichao also shared how the Manus team is thinking about the next steps for universal Agents and what the future differences between AI Agents and AGI might be.

"Nowadays, everyone seems to call everything an Agent. For example, a microphone, some might call it an 'environment-aware audio Agent,'" Ji Yichao joked.

He also outlined two main lines for expanding the capabilities of universal Agents: first, using multiple Agents to broaden the execution scale (such as deriving hundreds of parallel sub-Agents in large-scale research); second, opening up a larger "tool surface" for Agents, not limiting their capabilities to a few preset APIs, but allowing them to call upon the open-source ecosystem, install libraries, and even visualize and self-check through images.

Ji Yichao further mentioned that today's digital world is still built on a "for human use" paradigm—non-API web pages, CAPTCHAs, and gamified processes create significant friction, with bottlenecks more resembling ecological and institutional constraints rather than model intelligence.

This is also one of the reasons Manus participated in the Stripe event: both parties are promoting the completion of payments within the Agent, linking "research—decision—ordering/settlement" into a closed loop, using infrastructure collaboration to dissolve the world's friction.

Below are the highlights of the conversation, edited by Geek Park:

Q: Can you briefly introduce yourself to the audience? Your recent blog on "context engineering" is very inspiring, and I think it's a must-read for anyone developing AI Agents here. Every time I have lunch with engineers, they are always talking about it, so I can only sit somewhere else (laughs). But for those who may not be familiar with Manus, can you share your experience and vision?

A: Thank you, Paul. I'm glad to be here. Manus is building a universal AI Agent.

Many research institutions and companies are trying to create a brain—a large language model. But we believe that from a consumer's perspective, this is not ideal. AI should be able to take real actions and accomplish tasks, which is why we built Manus.

Our approach is to enable AI to use one of humanity's greatest inventions—universal computing. By giving AI a computer, it can do everything humans can do. Manus can truly complete tasks. For example, it can help you create presentations, plan a trip, or even manage social media—though I wouldn't recommend you actually do that.

Our users really love Manus. We launched Manus in March, and we have now achieved about $90 million in Revenue Run Rate (RRR), and we will soon surpass $100 million.

I think this is huge for a small startup like ours. But more importantly, it shows that AI Agents are no longer just a buzzword in research but are being applied and taking root in reality.

I can share a little story from our process of building Manus.

We actually drew a lot of inspiration from the application of Agent coding. For example, AI programming products like Cursor have already attracted a lot of attention.

As engineers, we naturally use Cursor. But what surprised us was that many non-engineer colleagues in the company were also using Cursor. Of course, they weren't writing software; they were using it for data visualization or even writing articles. They would ignore the code part on the left and just interact with the AI to get their work done.

This made us realize: we should generalize this approach and empower non-programmers. This is one use case for AI.

Q: We are hearing more and more discussions about AI Agents and AGI. Can you help us clearly distinguish between these two concepts? What do AI Agents and AGI mean to you and Manus?

A: We think this is a very good question.

Nowadays, almost everything is called an "Agent." For example, a microphone, some might say it is an "environment-aware audio Agent."

But at least we argue that an Agent should be a subset of applied AI. Let's take a step back and look at common categories of AI applications.

Most people are familiar with two categories: one is chatbots, like ChatGPT; the other is generative tools, like MidJourney or Sora. In these systems, there are usually only two roles: the user and the model. You interact with the model to get output. The difference with an Agent is that, in addition to the user and the model, it introduces a third key element—the environment.

The concept of "environment" varies depending on the type of Agent. For example, in design-oriented Agents, the environment might be a canvas or a piece of code; whereas in Manus, our goal is to have Agents appear in virtual machines or even across the entire internet. This way, Agents can observe the environment, decide what to do next, and take action to change the environment. This makes it very powerful.

For instance, in Manus, you can express a need, and it will open a browser, publish a webpage, or help you book a flight. I love this example because, although booking a flight sounds simple, it is actually AI directly changing the real world—the result is not just the model's output, but the ticket in your hand. AI is truly intervening in your world. This is what we call an Agent.

In simple terms, an Agent is an AI system that can interact with the environment on behalf of the user.

As for AGI, this term is often mentioned, and many equate it with superintelligence. We believe AGI is a system that can utilize the general capabilities of AI models to accomplish many tasks without special design.

We think "Agent coding" is actually a pathway to AGI. It is not a capability confined to a vertical domain; if you empower it to a computer, it can do almost anything on the computer. So for us, the condition for AGI is to build a sufficiently complete environment to enable this capability.

Q: In what scenarios is AI truly making an impact today? Where will it have an impact in the future? When will we see an iPhone moment?

A: In terms of Agents, if we look solely at model capabilities, the flagship models today are already quite impressive, almost "superhuman." They can outperform most of us in math competitions or logical reasoning.

But I believe models still resemble "brains in a bottle." To truly unleash their power, they must interact with the real world and reach reality. Unfortunately, this is where the problems begin.

For example, if you ask an AI to perform some transactional tasks, it is indeed very good at repetitive tasks. Products like Deep Research simply aggregate information and provide a result, and their output just appears there.

For instance, almost everything today is designed for human use, not just in the physical world but even in the digital world. For example, web tools are like mini-games, lacking APIs or standard interfaces. CAPTCHAs are everywhere, constantly intercepting Agents.

So I think AI performs well in closed, self-contained tasks, but once it involves the real world, it encounters obstacles.

When will we see an iPhone moment in the future? I think this is not a technical issue but more like an institutional limitation. This is not something that Agent startups like ours can solve alone.

I believe it requires a gradual transformation, necessitating the entire ecosystem to evolve together. It also requires companies like Stripe to make efforts at the infrastructure level. For example, we are integrating the new Stripe Agentic payment API. Everyone is working together.

Q: Can we talk specifically about some typical scenarios where users are using Manus? How are they using it? What power does this reflect?

A: Yes, although we come from the current generation of Agents, we have already seen many great use cases.

For example, we just moved to Singapore and needed to hire a real estate agent to help us find a place to live. A real human Agent (laughs).

Now these agents are already using Manus: they analyze the company's location and the areas where employees want to live based on client needs and generate corresponding recommendations.

I find this very interesting because it represents a kind of "long-tail demand." Generally speaking, there are no specific AI products designed for such specific scenarios, but since Manus is a universal Agent, it can meet these needs. We believe long-tail demand is very worthy of attention.

From a macro perspective, it may be long-tail, but for specific users, this is their daily work. Such scenarios are particularly valuable.

It's like today's search engine landscape. If you are just searching for common content, whether using Google or Bing, the quality of results is similar. So why do people choose one over the other? Perhaps because a particular search engine provided them with more suitable results at a specific moment. And if you search for very personalized or specialized content, the differences become even more apparent. So we believe the advantage of universal Agents lies here.

So how can we make it better? We have thought about this for a long time because we believe everything revolves around programming. If you hand a computer to AI, then its way of interacting with the environment is essentially through programming.

We think improvements can be made in two aspects. The first is scaling. But what if you could amplify the Agent's capabilities a hundredfold?

Recently, Manus has launched a new feature called Wide Research. The basic idea is to allow one Agent to derive hundreds of Agents to work together to complete tasks. You know, if you only let AI help you with small tasks, many times you can do it yourself. But if the task is very large, it is impossible for you to complete it alone, such as needing to conduct large-scale research. In this case, having hundreds of Agents working in parallel becomes very powerful.

Secondly, we also need to enable Agents to use computers more flexibly. For example, if you only set up preset tools for an AI Agent, then its action space is limited to those tools. But imagine if you are a programmer and have access to the entire open-source community's resources.

For instance, when you are 3D printing, directly modifying the model parameters can be difficult, but if you can find the right library on GitHub and install it directly, it can solve your problem. At Manus, we are optimizing for versatility and have proposed a concept called "network effects of tools."

There is an interesting example: many users are using Manus for data visualization. You know, in Asia, there can sometimes be issues, such as font errors when displaying Chinese in charts. Perhaps some professional users will write hard-coded rules, like which font to use when outputting Korean. But this approach can make the system increasingly rigid.

Our approach has been to add a very simple capability to the system: image viewing. The results were surprising—because today's models are already quite intelligent, they can check the generated visual images themselves, recognize errors, and then automatically correct them. We found that increasing the flexibility of tools can solve more problems than hard-coded rules.

Q: This is an exciting time. I'm really thrilled; I just wish I could be thirty again (laughs). Speaking of medical research, I know Manus is strong in this area as well. Have you observed some users using Manus for medical research?

A: Many people are already using Manus for research, not limited to medical research. We find this very interesting because there are indeed many so-called "deep research" products that help you collect a lot of information and do some analysis, but in the end, they only give you a markdown file or document. This is far from enough.

Many times, what researchers really need is a deliverable that can be directly presented to their boss or team. So we have strengthened the output of research results on Manus. For example, in medical research, there are often needs to generate formal reports, such as slide presentations. Therefore, we must optimize the AI's output capabilities to meet the needs of researchers. This is a "tooling" experience.

For instance, many users now first use Manus for research and then directly generate a website. You would find this completely different from traditional website building methods.

You should know that building a website itself is not difficult; the challenge is ensuring the reliability and accuracy of the data. So we believe it is best to complete the entire process in one conversation, within a shared context. This way, your research and insights can seamlessly transform into final results. This is what we are doing in Manus.

Q: Many countries are discussing a topic: the future of humanity and economic impact in the AI era. What do you think about job displacement? What new job opportunities will emerge?

A: Our friends and investors often ask us this question. When we launched Manus, we initially thought that if we could build such an Agent, it would help people save a lot of time and make it easier for them to earn money.

But in reality, we found that this vision has not been fully realized. Through extensive user research, we discovered that after using it, users actually end up working more. Because they become more efficient, they can actually do more of what they are already good at. This is the first point.

Secondly, we believe Manus has opened up a whole new space. We have been discussing virtual machines and cloud computing. We think Manus is playing a role as a "personal cloud computing platform." For example, cloud computing has existed for decades, but it has mostly been the privilege of engineers, who are the only ones able to harness the power of the cloud through programming. Ordinary knowledge workers cannot use it.

But now, with AI Agents like Manus, people can issue commands in natural language and let AI execute them. This unlocks a whole new level of productivity. This is what we bring.

Finally, regarding "replacement," I think it is actually very difficult. For example, real estate agents use Manus to complete their daily work. But you know, AI can never replace the way agents communicate with clients. We are an AI company, and even the script for Manus's launch video was written by Manus, but the person appearing in the video is still me, because this is about trust. And trust cannot be completely entrusted to AI.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。