There is a brutal irony in the story of Manus.
At the end of 2025, Manus looked like one of the most important AI-agent startups in the world. Meta agreed to acquire the Singapore-based company, which had Chinese roots, for more than $2 billion. Manus had built its reputation around something that was still relatively novel at the time: an AI system that could move beyond answering questions and actually execute multi-step tasks on a computer.
Then geopolitics intervened.
In April 2026, Chinese regulators ordered Meta to unwind the transaction, citing national-security concerns over the transfer of Chinese-origin AI technology and talent to a U.S. company. Meta subsequently separated its operations from Manus, including cutting off data sharing and internal access.
The result was extraordinary.
Meta had committed more than $2 billion to acquire one of the hottest agent companies in the world — and within months, the deal was being dismantled.
By June, Reuters reported that Manus's early Chinese investors were considering buying the company back from Meta for approximately the same $2 billion valuation. By August, Manus was transitioning back to independent operations, with some user data generated after Meta's acquisition subject to deletion as part of the regulatory separation. Manus explicitly said the deletion was not the result of a security breach, but a consequence of the separation and regulatory requirements.
And then came Muse.
Meta didn't leave the agent race behind
On September 8, Meta launched Muse, its personal AI agent.
The difference is scale.
Manus had to convince users to come to Manus.
Meta already has billions of people inside its ecosystem.
Muse is designed to operate through Meta's existing products and connected services. Meta says it can handle tasks such as email, travel planning and other multi-step activities, while running inside a dedicated Muse Secure VM — essentially a separate virtual computer in the cloud for each user's agent.
That is an important architectural decision.
Instead of merely giving an AI model access to a collection of APIs, Meta is attempting to give the agent its own controlled computing environment.
The VM contains the agent, browser and associated data. Meta says a separate security system called Sentinel controls whether Muse can reach the internet, while sensitive actions can require user approval and activity is recorded through an audit trail.
That gives Meta something Manus always had to fight for:
infrastructure at enormous scale.
And Meta is already pushing Muse beyond a standalone chatbot.
At Meta Connect, Zuckerberg demonstrated Muse across Meta's ecosystem, including smart glasses. Meta has also introduced a physical Muse device and is expanding Muse into small-business workflows.
That changes the competitive equation.
The most painful part for Manus
Manus 2.0 arrived on September 28.
And ironically, it is not a minor update.
Manus describes it as a new architecture built around an agent harness called Cascade. The company says one tested configuration used 23.2% fewer tokens, completed tasks 28.2% faster and cost 32% less to operate than its previous system. It has also introduced Cloud Computer, Automations, Manus Studio and a new personal-agent product called Cue.
So Manus is not technologically dead.
Quite the opposite.
It is still trying to innovate.
The problem is that innovation is no longer enough.
The market has changed underneath it.
When Manus became famous, being able to give an AI agent a computer and tell it to accomplish a complicated objective felt revolutionary.
Now that architecture is becoming a baseline expectation.
Meta has a dedicated VM.
OpenAI is developing its own autonomous agents.
Other startups are building persistent agents with their own environments, identities and tools.
The competition is moving from:
“Can an AI agent do things?”
to:
“Which agent will become the user's default digital worker?”
That is a much harder battle.
Did Meta simply take Manus and rename it Muse?
This is where the viral version of the story needs to be separated from the documented facts.
There is no reliable evidence showing that Meta simply took Manus's intellectual property, put a new name on it and released Muse.
Meta says Muse was built from the ground up.
At the same time, Meta owned Manus for several months, so it is reasonable to ask what knowledge, engineering experience and lessons Meta gained during that period.
But that is different from proving that Muse is simply Manus under another name.
The more interesting story is actually bigger than that.
Meta paid billions for access to a leading agent company.
China prevented the transaction from becoming a normal acquisition.
Meta then proceeded to build its own consumer-agent ecosystem anyway.
That is a story about how quickly the strategic value of AI startups can change when frontier technology collides with geopolitics.
Manus also faces a trust problem
Manus says it has substantial security controls and currently lists SOC 2 and ISO certifications through its Trust Center. It also says its agents operate inside isolated cloud virtual machines.
So describing Manus simply as “unsafe” would be unfair.
But AI agents inherently create a different security problem from ordinary chatbots.
An agent doesn't merely tell you what to do.
You give it permission to do something.
That means it may interact with your files, browser, accounts, messages, credentials and external services.
The industry is now discovering how difficult that becomes at scale.
Interestingly, Muse itself has already experienced security and privacy problems. A vulnerability was reported and patched, and a recent incident involved Muse sharing a user's home address with a Facebook Marketplace buyer after the user had granted broad permissions. Meta acknowledged the incident and said it was investigating and improving the permission interface.
That matters because it destroys the simplistic idea that Meta automatically solved the agent-security problem.
It didn't.
Nobody has.
The race is still happening in real time.
Manus's real problem isn't that Muse is better
The real problem is distribution.
A startup can build a technically impressive agent.
But Meta can place an agent inside the products people already use.
That is a completely different advantage.
Meta can connect Muse to Facebook, Instagram, WhatsApp, smart glasses and eventually other hardware.
On September 29, Meta announced Muse for small businesses, with integrations including Canva, Dropbox, Figma, Notion, Shopify, Slack, Stripe, Zoom and others.
That is not simply an AI model competing with another AI model.
It is an ecosystem competing with a standalone product.
And that is where Manus has to find a new identity.
Manus 2.0 may actually be its escape route
The irony is that Manus 2.0 is moving in a direction that could make the company more interesting, not less.
Its new architecture is designed around persistent projects rather than one-off prompts.
Cloud Computer gives projects their own computing environment.
Automations allow work to happen when events occur rather than only when a user manually starts a task.
And Cue pushes Manus toward personal agents with their own digital identity and resources.
That could give Manus a different battlefield.
Instead of trying to become “the other Meta Muse,” Manus could become infrastructure for people who want autonomous digital workers that are portable across ecosystems.
That is a harder proposition to sell.
But it is potentially much more defensible.
The Manus story is really a warning
The most fascinating part of the Manus saga isn't that a startup supposedly lost to Meta.
It is that AI is moving so quickly that a $2 billion acquisition can become strategically complicated within months.
In December, Meta wanted Manus.
Four months later, Beijing had forced the deal toward reversal.
By the summer, the companies were separating their systems and data.
By September, Meta was aggressively pushing its own agent into the consumer market.
And on September 28, Manus launched its biggest architectural overhaul yet.
That is an extraordinary amount of change in less than a year.
Manus isn't necessarily finished.
But it has lost something extremely valuable:
the ability to define the AI-agent conversation by itself.
Muse now has Meta's distribution, capital, hardware ambitions and ecosystem behind it.
Manus has to prove that it can offer something Meta cannot simply reproduce at scale.
That is the real test.
And if Manus succeeds, its story won't be remembered as the startup Meta bought and lost.
It will be remembered as the company that survived one of the strangest acquisition reversals in the history of AI — and reinvented itself after the world's biggest technology companies caught up.
Comments
Post a Comment