Manus AI: The Smart Agent That Redefined Digital Assistants and Attracted Meta’s $2B Attention

On 5 March 2025, the “smart assistant” was no longer just a chatbot that answers questions; it had evolved into a digital entity with hands to act. Manus AI, emerging from the labs of the Chinese start‑up “Monica”, became the first general‑purpose autonomous AI agent capable of turning ideas into tangible actions. Within just four hours of its announcement, visits to its website exceeded ten million, and obtaining an invitation code became a dream whose price in the black market exceeded ten thousand dollars. What started as a massive media buzz ended with one of the largest acquisitions in the sector’s history when Meta absorbed it into its empire by the end of the same year.

From an Idea to Global Frenzy: The Engineer Behind the Phenomenon

Behind Manus AI stands Chinese founder Xiao Hong, a graduate of Huazhong University of Science and Technology and founder of the previous company “Yi Ying” that served over 200 million users within the “WeChat” ecosystem. After selling his first project, Hong launched “Monica” in 2022 as a browser extension that combined capabilities from several AI models. By 2024, its users exceeded ten million, but his ambition went beyond the limits of a “smart assistant” toward building a real “smart agent”. Thus Manus AI was born, described by its developers as “the brain that grew two hands”, to become the world’s first general‑purpose autonomous AI agent.

The Multi‑Task Machine: A Technical Architecture That Gathers “the Best” Not “the Latest”

Unlike giant foundation models that rely on a unified architecture, Manus adopted a composite architecture based on a “fusion” strategy. It did not depend on a single language model built from scratch; instead, it employed powerful external models such as “Claude” from Anthropic and “Qwen” models from Alibaba, which were fine‑tuned and trained specifically for specific roles inside a multi‑agent system.

The secret was not the power of a single model, but Manus’s ability to coordinate a set of specialised sub‑models within an isolated virtual environment. Upon receiving any task, the agent creates a “virtual computer” dedicated to that task in the cloud. This environment is equipped with a full operating system, a web browser, software‑development tools, and databases, allowing it to write and execute code, browse the Internet, edit files, and create complete websites and applications completely independently. This architecture, based on a “zero‑trust” model, gave the agent broad permissions inside its sandbox without risks to user data.

Capabilities That Redefined “Autonomous Work”

Manus differs fundamentally from traditional smart assistants by the principle of autonomous execution – it does not merely offer advice; it takes over the task from start to finish without continuous human intervention. Inside the virtual sandbox, the agent can:

• Browse the Internet and conduct in‑depth research
• Write and execute Python code for data‑analysis operations
• Create full web applications from front‑end to back‑end
• Analyse CSV and Excel files, generate charts, and export results to PDF or presentation files

To simplify use, Manus provided a library of 32 ready‑made templates covering five key domains: business intelligence (e.g., SWOT‑analysis generator, start‑up‑idea validation), marketing (e.g., market‑research tool, YouTube influencer researcher), creativity and design (e.g., video generator, presentation maker). But the most outstanding achievement was the “horizontal search” feature, designed to work horizontally by executing multiple parallel tasks simultaneously, unlike the “vertical search” adopted by OpenAI that delves into a single topic. In practical tests, Manus could compare a hundred sneakers at once or create fifty advertisement posters with different styles in a single session.

Multi‑Agent Architecture: The Specialised Digital Team

The true secret of Manus’s power lies in its multi‑agent architecture. Instead of relying on one giant AI model, the agent uses an integrated team of specialised agents that collaborate:

• The Planner Agent understands the task and breaks it into small steps
• The Executor Agent actually executes these steps by writing code or browsing the web
• The Reviewer Agent checks the results and verifies they match the original goal before delivery

The system relies on the CodeAct engine, which uses the Python programming language as a “universal communication language” with tools. Instead of complex JSON formats, the executor writes Python code to interact with the browser or analyse data, giving it greater flexibility and efficiency. Manus also works asynchronously – the user can send the task and close the app, while the agent continues working in the cloud and sends a notification when finished, a key feature that differs from tools like Cursor that require continuous monitoring.

Record‑Breaking Numbers and Stunning Financial Achievements

Manus achieved record results in the GAIA benchmark, the most important standard for measuring AI’s ability to solve real‑world problems. Manus’s overall score reached 86.5 %, significantly ahead of its competitors – the previous best achievement by OpenAI DeepResearch was only 65 %. By task‑difficulty level, it scored 86.5 % in basic tasks, 70.1 % in intermediate tasks, and 57.7 % in complex tasks.

Financially, Manus became one of the fastest‑profit‑generating AI products in history. Within just eight months of its launch, its annual recurring revenue exceeded the $100‑million barrier, while its revenue‑consumption rate reached $125 million. These numbers would not have been possible without the tangible value the agent delivers – a single Manus user consumes computational‑token quantities equivalent to 1,500× what a traditional chatbot user consumes, reflecting the complexity of the tasks it performs.

Launch Buzz and Unprecedented Spread

At its launch in March 2025, visits to the Manus website reached ten million within just four hours. Hundreds of thousands were placed on waiting lists for an invitation code, and the price of a single code on the digital black market reached the equivalent of $13,900 (about 100,000 Chinese yuan). According to company statistics at the beginning of December 2025, Manus processed over 147 trillion computational tokens and created over eighty million virtual computers to serve millions of users worldwide.

But this growth did not continue in a straight line. After peaking at 23.76 million visits in March 2025, visitor numbers gradually declined to 17.56 million in August 2025, and this downward trend continued in 2026, raising questions about the platform’s ability to retain users long‑term.

Technical Controversy: The “Wrapper” Allegation and Practical Limitations

Manus received widespread criticism because it does not rely on an AI model developed in‑house, but uses external models such as Claude 3.5/3.7 Sonnet from Anthropic and Qwen from Alibaba. Some observers described it as nothing more than a “skilfully crafted wrapper” and not a fundamental innovation. However, this argument overlooks the fact that Manus’s real value lies not in the language model, but in its superior ability to coordinate these different models and tools within one seamless working environment.

On a practical level, tests have shown that Manus has not yet reached a state of “full reliability”. In a comparative test conducted by the AGI‑Eval community, Manus’s best score was 2.20 out of 5.00, meaning it is still in the “partial‑use” stage, not a total‑reliance solution. Moreover, high operational cost poses a challenge – due to its complex working mechanism, Manus consumes enormous amounts of computational resources, making it an expensive solution compared to alternatives.

The Acquisition That Shook Silicon Valley: Why Did Meta Want Manus?

After months of speculation, Meta officially announced on 30 December 2025 its complete acquisition of Manus AI and its parent company “Butterfly Effect”. Estimates of the deal’s value ranged between $2–3 billion, with some reports suggesting the figure could exceed $20 billion, making it one of the largest acquisitions in Meta’s history after the WhatsApp purchase in 2014 and its investment in Scale AI.

The strategic motivation was clear: after Meta spent tens of billions on data centres and developing foundational models like “Llama”, Manus represented the missing “execution layer”. It was not just a promising research project, but a real business generating huge profits with a high customer‑retention rate. Manus proved that users are willing to pay monthly subscriptions of up to $199 for a digital agent that can execute tasks on their behalf.

Under the deal, Xiao Hong moved to work as Vice‑President at Meta, responsible for smart‑agent‑product strategy, reporting directly to Mark Zuckerberg, who publicly admitted he was a loyal Manus user. The agreement also stipulated that Manus would continue to operate as an independent entity from its new headquarters in Singapore, with all its operations and services within China terminated, and ensuring no ongoing Chinese ownership of the platform after the deal’s completion.

Post‑Acquisition: Strategic Integration and the Future of Smart Agents

Only a few months after the deal, signs of strategic integration began to appear. On 16 March 2026, Manus launched a new desktop application called “My Computer” for Mac and Windows systems. This application marked a qualitative leap in the agent’s capabilities, enabling it to read and edit files on the user’s local machine, run and control local applications, and execute complex software projects without uploading any files to the cloud.

This application represents Meta’s direct response to the rise of open‑source tools like “OpenClaw”. While those tools are free and customisable, Meta bet on offering a polished, reliable commercial alternative that works consistently out‑of‑the‑box, backed by Meta’s own models. At the same time, reports surfaced that Meta is working on integrating Manus’s capabilities with its internal “Avocado” model, indicating a future where AI will be able to operate seamlessly between the cloud and personal devices.

Challenges of Digital Agency: Limits of Trust and Responsibility

With the acquisition chapter closed and a new era under Meta’s umbrella beginning, Manus stands at a crossroads that raises more questions than it answers. Integrating Manus’s technologies across Facebook, Instagram, and WhatsApp platforms opens the door to billions of users, but at the same time raises complex questions about the limits of digital “agency”. Allowing a smart entity to access personal accounts and execute tasks on behalf of users raises profound issues concerning legal and moral responsibility for its actions.

More importantly, can agents like Manus one day overcome the “reliability” hurdle that still troubles generative AI models? However advanced an agent may be, it remains prone to errors and hallucinations, and the only difference is that its mistake is no longer limited to providing incorrect information—it could extend to taking actual actions in the digital or financial world on the user’s behalf. In this new landscape, the question is no longer “What can AI do?”, but “How far do we trust it to do that on its own?”