Today, Manus officially announced its joining of Meta.

According to LatePost, the acquisition deal is worth billions of dollars, marking Meta's third-largest acquisition since its inception, following WhatsApp and Scale AI.

This was a classic "blitzkrieg" move. Before Meta's intervention, Manus was seeking a new round of financing at a valuation of $2 billion. Meta sealed this massive deal in just over ten days. Upon completion, Manus founder Xiao Hong will serve as Meta's Vice President, and its core technical team will be fully integrated into Meta's AI department.

Meta's urgency to acquire this team stems not just from the code but from their unique approach to AI.

Over the past year, the Manus team has been like explorers on the edge of a new continent, sharing many insightful perspectives ranging from product philosophy to competitive strategy.

Here, we've compiled eight of the most valuable, even somewhat "counterintuitive" viewpoints.

Why Manus? The answer to this billion-dollar question was already written in their every past expression.

/ 01 /

From Browser to Manus

The birth of Manus actually stemmed from an expensive "trial and error."

Prior to this, the team had developed an AI browser, similar in form to the later Arc Dia and Perplexity Comet, with a high degree of maturity, even reaching the point of being ready for launch in a week.

For many, an AI browser riding the wave was a better choice. However, at the last moment before release, the project was halted.

The reason stemmed from a fascinating phenomenon observed by the team: AI was exceptionally skilled at manipulating browsers, to the point of "fighting" with humans.

This completely subverted the logic of an AI browser. It felt like your intern showing up without a laptop and insisting on using yours to get work done.

Thus, the first conclusion emerged: AI is adept at using computers, but it shouldn't use yours; it should have its own.

This directly shaped Manus's later form: not just a chatbox, but an AI equipped with a virtual machine in the cloud. The Agent must have an independent operating environment in the cloud, without interfering with the user.

The second key inspiration came from Cursor.

After Cursor emerged in June last year, Manus partner Zhang Tao noticed that colleagues and even family members who couldn't code started using Cursor to solve everyday problems, like his wife converting a video file from MP4 to MP3.

This made them realize something: the benefits of AI coding were, for the first time, reaching non-engineers.

Many, upon seeing Cursor, wanted to create "even stronger engineer tools." Their judgment was the opposite: the world already had too many fancy tools for engineers. Engineers didn't need another fancy tool. The real opportunity was to democratize AI coding, to give programming—a way to mobilize the world—to every ordinary person.

Thus, the second conclusion emerged: the potential of AI coding should be democratized, allowing non-coders to enjoy its benefits.

Combining these two insights, the initial form of Manus was finally outlined:

AI should have its own computer (cloud environment, virtual machine, controllable toolchain), enabling ordinary people to use AI for coding and more complex tasks.

/ 02 /

Be Hao123, or Be Baidu

From the outset, the Manus team made a crucial, somewhat counterintuitive decision: to insist on creating a "general-purpose" Agent.

Xiao Hong used a very apt analogy:

Vertical Agents are like Hao123: a series of preset slots by developers. You stack functions like links, and users can only navigate within the predefined boundaries. This model seems precise but quickly hits an evolutionary ceiling.

General-purpose Agents are like Baidu: start with a "box" that understands everything. Let users input freely, and when millions start searching for "how to book a ticket," optimize a "ticket booking card" in response.

This insight was highly forward-looking at the time. The industry was filled with voices for "vertical Agents" (e.g., travel planning Agents), but the Manus team was certain: Agents must first be general-purpose to truly survive. Otherwise, the high customer acquisition costs in vertical scenarios would crush the entire business model.

The seed of this insight was actually sown during the development of Monica.

Once, Xiao Hong discussed with Youzan founder Bai Ya about taking Monica's feature integration to the extreme. Bai Ya responded with a highly inspiring remark: "Red, being extreme isn't enough; being personalized is. If you're extreme, you're Hao123; if you're personalized, you're Google."

Later, the Manus team spent considerable time studying this statement and ultimately applied it to Manus.

/ 03 /

Model Capabilities Evolve, So Must the "Shell"

In a March interview this year, Xiao Hong offered a profound industry observation: as models evolve, so must the "shells" that house them.

In his view, each leap in model capabilities requires a new product form to unleash its value.

· Jasper emerged earliest, with fill-in-the-blank interactions;

· ChatGPT introduced the chatbox;

· Monica featured a sidebar with built-in context;

· Cursor became an editor capable of directly rewriting code;

· Manus is an Agent with an independent cloud environment;

This is the "Andy-Bill's Law" of the AI era: the benefits provided by hardware (models) are consumed by software (product forms). And Manus is the "shell" attempting to consume the benefits of the new generation of models.

In shell design, the product's interactive interface is considered key to user acceptance.

When discussing UI, the team critically examined Devin's interface. Devin cluttered the screen with Planner, Shell, and Browser, imposing a huge cognitive burden on non-technical users.

Manus proposed the concepts of "progressive disclosure" and "operating system metaphor":

Progressive Disclosure: The default interface should be extremely simple, like Google's search box. As tasks unfold, the necessary tools (Shell, Browser) emerge like "floating to the surface."

Operating System Metaphor: Design different core functions (e.g., browser, spreadsheets, document editors) as independent, equal "first-level applications" rather than chaotically nested. Users can switch between these "applications" just like in Windows or macOS, providing a clear, scalable framework for future expansions.

/ 04 /

Philosophy of Trust: Less Structure, More Intelligence

In March, amid the prevalence of Workflow, Manus chose another path: Zero Predefined Workflow.

Product managers often view technological progress statically, only wanting to "solve user problems." However, the Manus team prefers a more developmental perspective, considering questions like, "The next generation of models might solve this."

When you truly return intelligence to the model, magic happens. AI's ability to handle long-tail tasks often surpasses that of human-designed processes. This is a philosophy of trust: rather than trying to control every step of AI, believe in the power of Intelligence itself.

/ 05 /

Agent Competition Is System Competition

All Agent companies face the ultimate question: What if OpenAI enters the game?

The Manus team believes that model companies entering the field are not guaranteed to win.

The direct reason is that model companies can only use their own models, which is both an advantage and a constraint.

Manus can call upon the most suitable model for each task: Gemini for search, GPT-5 for reasoning, and Claude for coding. This flexibility is something ChatGPT's product managers cannot possess.

A deeper reason is their belief that Agent competition is essentially system-level competition.

The biggest difference between Agents and AGI is that Agents introduce a third key element besides users and models: the environment.

The concept of "environment" varies with the type of Agent. For design-oriented Agents, the environment might be a canvas or a code segment; for Manus, the goal is to have Agents operate in virtual machines or even the entire internet.

This means Agents compete not just on model capabilities but on how to construct the surrounding environment and prepare the toolkit.

They believe there's a vast amount of engineering work to be done here.

For instance, the biggest shortcoming (weakness) of current Agent products (e.g., Devin) lies in their "one-time" session mechanism. Each task operates in a brand-new, sterile environment, leading to redundant work and a poor user experience.

In the Manus team's view, achieving persistent login states is the cornerstone of true "agency." Agents must be able to maintain login states across various websites, avoiding the need for manual user intervention each time.

This way, users only need to log in once, and the Agent can operate on their behalf long-term (long-term).

/ 06 /

Visibility Matters More Than Being First

Manus's success lies in delivering the right product at the right time.

Manus initially benefited from organic traffic but remained sober (soberly aware) of its pitfalls.

Organic traffic attracts "innovators" and "early adopters" who are forgiving, curious, and eager to try new things. However, crossing the chasm into the mainstream market follows entirely different rules. The mainstream doesn't care if you're the "first general-purpose Agent"; they only care, "What can you do for me?"

You must then consider, on which media and channels, and in what ways, to convey the most effective information to them. The key to marketing is information transmission efficiency, not simply spending money.

/ 07 /

Hands-On Product Experience

In interviews with the Manus team, a crucial point mentioned is that the core team always personally oversees the most critical aspects of the product experience.

The initial decision to pursue this path stemmed from a consensus among core members. However, before the product materialized, this consensus only existed in our minds. If more people were involved in execution without sufficient context and intuition, it would easily go astray.

Thus, from writing the first line of code for Manus to the first forty days of development, the project involved only five people.

The advantage of this approach was high alignment and extremely efficient communication. Whether it was writing each prompt, designing the overall technical framework of the virtual machine, or refining product interaction details, everything remained within the scope of these five people, ensuring smooth progress.

/ 08 /

Think in Game Theory, Not Just Logical Reasoning

For AI entrepreneurs, Xiao Hong offers an ultimate piece of advice: think in terms of game theory, not just logical reasoning.

Logical reasoning is linear: Google has the strongest technology, so Google will win.

Game-theoretic thinking is dynamic: because of OpenAI's existence, Google's strategy was forced to change; because of your presence, the decision-making logic of giants will also change.

Entrepreneurship is not about solving problems in a vacuum but about finding an ecological niche in a multi-player game ecosystem.

For instance, if a powerful third party had created ChatGPT back then, OpenAI might have chosen to become a pure Platform Company.

Founders must possess this god-like perspective: see variables, anticipate gameplay, and then make decisive moves.

By Lang Lang, Lin Bai