A week of Manus

Seven days after launch, the new all-in-one agent Manus AI still commands attention and headlines. The company, developed by Chinese startup Butterfly Effect, claims 2 million people are on the waitlist for its general purpose agent. Adding to the hype, Manus is still only accessible by invitation, though invite codes can be bought on the resale market and on social media for thousands of dollars.

And Manus’s ambition might be bigger than generating hype and comparisons to another Chinese model, DeepSeek, which sent shockwaves through the tech and financial worlds earlier this year. Two days ago, Manus AI announced a strategic partnership with Alibaba’s Qwen team, a move that could help the startup respond to the surge in traffic and expand its user base, particularly in China.

But whether Manus amounts to more than a splashy model demo that went viral remains to be seen. “A use case, a demo, a POC [proof of concept] is very different from what can you integrate in an enterprise architecture, right?” says Vyoma Gajjar, an AI Technical Solution Architect at IBM, in a recent episode of Mixture of Experts.

To scale real-world applications, we need “specialized, cost-efficient agents that are specially configured for enterprise use,” says IBM’s Maryam Ashoori, Director of Product Management for watsonx.ai in an interview with IBM Think.

Manus dazzled many users as it could quickly execute a wide range of complicated tasks from analyzing stocks to scheduling trips to evaluating insurance, all without any human input besides the initial prompt.

“Manus is a really nice automation experience where you can chat with the model, and the model has access to 29 tools, and it orchestrates your requests and it gives you really nice progressive feedback of how it's performing that task,” explains Chris Hay, Distinguished Engineer at IBM, in an interview with Think.

Recent months have been marked by the release of many new AI tools and features, underlining a growing appetite for agents supporting users in various tasks (think OpenAI’s Operator, or the various Deep Research features that were recently added to many LLMs).

“Manus is this nice mix of deep search, deep research and browsing. And it also has access to other tools,” says Hay.

Manus’s real innovation might be to bring together many interesting tools (an editor, a browser, a terminal) in a user-friendly experience, which, Hay believes, echoed the first public release of ChatGPT over two years ago. At that time—and for the first time—interacting with AI became accessible to anyone with an internet connection, no computer science degree required.

Manus uses several AI models, notably Anthropic’s Claude, Qwen and Browser Use software that helps AI agents access websites. Browser Use is enjoying the success halo of Manus’s breakout week by going viral itself.

But will anyone be talking about Manus next month? Hay predicts, “In a week, someone will recreate Manus only with open source, which I think is cool. OpenAI will respond, and everyone will have an equivalent type of experience.”

While Manus has captured consumer interest, enterprise adoption of this type of autonomous general agent might be a longer road, if it makes sense at all. The LLMs powering agents that have drawn so much attention in the last 12 months are mostly trained on publicly available data. In contrast, only 1% of enterprise data is currently used in large language models.

The 99% of data that is not captured represents a gold mine for those companies looking to use LLMs and agents to unlock the value in that data. Enterprise AI agents, however, are generally used differently than personal AI agents. For instance, individuals might use an agent built by a company such as Manus AI or OpenAI to schedule a flight or review their resume.

Enterprises, however, often seek an agent system built with either closed or open-source tools that has been customized for their specific business needs to be as cost efficient as possible (think Salesforce’s Agentforce or IBM's watsonx Orchestrate). Companies also have a tremendous amount of data they do not want to move off premises for security reasons. These enterprise agents are typically powered by smaller fit-for-purpose language models, such as IBM’s Granite series, that keep necessary data on site.

Whether Manus has enterprise AI ambitions is unclear, however, its newly announced partner, Alibaba, certainly does. In January, Alibaba announced a new assistant, Qwen 2.5 Max, which runs on the mixture of experts (MoE) architecture, an efficient approach that often appeals to enterprises.

With an MoE architecture, the model only activates the specific sub-networks needed for a given task rather than activating the entire neural network. Consequently, the MoE architecture greatly reduces computation costs during pre-training and achieves faster performance during inference time.

Whether true AI autonomy is achieved with a personal AI agent or an enterprise AI agent remains to be seen. Manus, meanwhile, continues to spark debate.

“Manus is definitely shaking things up here a bit. But there are also a lot of skeptics in the AI community,” says Kaoutar El Maghraoui, a Principal Research Scientist and Manager at IBM in a recent Mixture of Experts episode. “The big question here is, can Manus really redefine AI autonomy? Or is it just another step in an ongoing AI race between East and West?"