Manus: The AI Agent That Went From Launch to $2 Billion Meta Acquisition in Nine Months

In March 2025, a relatively unknown Chinese startup called Butterfly Effect (better known as Monica) dropped an AI agent called Manus. Within hours, AI Twitter lost its collective mind. Within months, it was pulling in $100 million in annualized revenue. By December, Meta had written a $2 billion check to buy the whole thing.

That trajectory — from “wait, what is this?” to a mega-acquisition by one of the largest companies on Earth — tells you something about where the AI industry is headed. And it’s not just about chatbots getting smarter. It’s about agents that actually do things.

What Manus Actually Is

Forget the marketing copy. Here’s what Manus does at a mechanical level: you give it a task in plain English, and it goes away and does it. Not “generates a response about how to do it.” Not “writes you a plan.” It opens a browser, navigates websites, writes and executes code, reads documents, creates spreadsheets, builds dashboards, and comes back with finished work.

The key difference from a chatbot is the word autonomously. You don’t sit there prompting it step by step. You say “research the competitive landscape for electric vehicle charging stations in the Northeast US and build me a comparison spreadsheet,” and it actually does that. It browses the web, pulls data, structures it, and delivers a file. You can close your laptop and come back later.

Under the hood, Manus doesn’t use its own AI model. It orchestrates other models — primarily Anthropic’s Claude — through a multi-agent architecture. There’s a planner that breaks your request into subtasks, and executor agents that carry them out using 27 different tools. It’s less “a new AI” and more “a very sophisticated wrapper that makes existing AI actually useful for real work.”

The team was upfront about this. They never pretended to have a frontier model. They built the orchestration layer — the part that decides what to do, when, and how to recover when things break.

The Virtual Computer Trick

The most interesting technical detail is where Manus runs. Each task gets its own virtual computer — a full Ubuntu Linux environment with internet access, a browser, a filesystem, and sudo privileges. These aren’t Docker containers. The team tried Docker first and found it too limited. They needed a real operating system so the agent could install packages, launch web servers, and behave like an actual power user sitting at a workstation.

They ended up using E2B, which runs Firecracker microVMs — the same lightweight VM technology that powers AWS Lambda. Each sandbox spins up in about 150 milliseconds, runs for as long as the task needs (hours, if necessary), and gets destroyed when it’s done. The agent can pause and resume, persist data for up to 14 days, and even expose locally-running web servers to the internet.

This matters because it’s the difference between “AI that talks about doing things” and “AI that has a computer and does things.” Manus can install Python packages mid-task, open a Chromium browser and click through web apps, write files to disk, and run shell commands. It operates more like a remote freelancer than a chatbot.

The Hype, Then the Backlash

Manus launched to genuine excitement. Hugging Face’s Victor Mustar called it “the most impressive AI tool I’ve ever encountered.” It topped the GAIA benchmark — a test of real-world AI agent capabilities — scoring 86.5% on basic tasks, 70.1% on intermediate, and 57.7% on hard ones. Those numbers beat OpenAI’s Deep Research system across the board.

Then reality set in.

Users reported constant server overload errors. The platform crashed during live tasks. Someone tried to get it to order a sandwich and it fell apart. It generated broken links for flight bookings. It sometimes cited sources that didn’t exist. The invite-only access model and server capacity issues led critics to accuse the team of “hunger marketing” — building artificial scarcity to fuel hype.

Opinions split into predictable camps. Enthusiasts pointed to the benchmark scores and the genuine novelty of a fully autonomous agent. Skeptics pointed to the reliability problems and the fact that Manus was essentially a UI layer on top of Claude. The measured crowd said both things were true — it was genuinely impressive and genuinely not ready.

The most substantive criticism: Manus didn’t build its own foundation model. It depended entirely on external AI providers, primarily Anthropic. That’s a strategic vulnerability. If Claude’s API pricing changes, or Anthropic decides to build its own agent product (which, they have), Manus is exposed.

The Meta Deal

On December 29, 2025, Meta announced it was acquiring Manus for a reported $2 billion. The company had relocated from China to Singapore, and Meta specified that “there will be no continuing Chinese ownership interests” post-acquisition — a necessary political move given the regulatory climate.

The strategic logic for Meta is straightforward. Meta has Llama, its open-source model family. What Meta didn’t have was an execution layer — the agent infrastructure that turns a language model into something that can actually complete tasks. Manus gives them that. The plan is to integrate Manus’s agent capabilities into Meta AI, the assistant that lives across WhatsApp, Instagram, Facebook, and Ray-Ban Meta glasses.

Think about what that means: an autonomous agent that can browse the web, write code, and manage complex workflows, embedded into apps used by billions of people. Whether that’s exciting or terrifying probably depends on how much you trust Meta with that kind of capability.

The acquisition was Meta’s fifth pure AI deal of 2025, part of a year where the company spent at least $70 billion on AI infrastructure. They’re not dabbling. They’re building something.

The Customer Exodus

Not everyone was thrilled. After the acquisition, some existing Manus customers walked.

Seth Dobrin, CEO of Arya Labs, put it bluntly: he trusted Manus’s transparent terms of service, but he doesn’t extend that trust to Meta. His company stopped using the platform entirely. He’s not alone. The concern isn’t about the technology degrading — it’s about data governance, privacy policies, and who ultimately controls the agent infrastructure your business depends on.

This is a pattern worth watching. Every major acquisition in AI triggers the same calculation: is the product going to get better because it has more resources, or worse because it now serves a different master? With Meta’s track record on data privacy, the skepticism isn’t unfounded.

What This Tells Us About Where Agents Are Going

Manus matters less as a specific product and more as a proof point. It demonstrated three things:

Orchestration can be the product. You don’t need to build a frontier model to build a valuable AI company. Manus got to $100 million in revenue by being good at the plumbing — task decomposition, tool use, error recovery, sandbox management. The models were commodities. The coordination was the value.

Agents need real computing environments. The sandbox-per-task approach — giving each agent its own virtual computer — is probably the future of AI agent infrastructure. Chatbots live in text boxes. Agents need operating systems. E2B, the company providing Manus’s sandboxes, raised $21 million off the back of this trend. Expect more infrastructure companies in this space.

The trust problem is real. The Meta acquisition backlash showed that enterprise customers care deeply about who controls the AI agent they’re feeding sensitive data to. Autonomous agents that browse the web, execute code, and handle business workflows create a much larger trust surface than a chatbot that generates text. The companies that solve the governance and transparency problem will win enterprise adoption. The ones that don’t will keep losing customers every time an acquisition press release drops.

The Bigger Picture

A year ago, “AI agent” was a buzzword people threw around at conferences. Today, one sold for $2 billion, and the acquirer is integrating it into products used by three billion people.

The gap between “AI that responds” and “AI that acts” is closing fast. Manus wasn’t the first agent, and it won’t be the last. But its arc — from a scrappy Chinese startup to a Meta acquisition in under a year — compressed an entire industry cycle into a few months. That speed is the story.

Whether Meta turns Manus into something transformative or buries it in the organizational chart remains to be seen. But the blueprint is out there now: take existing models, give them real computers, teach them to use tools, and let them work autonomously. The next Manus is probably already being built.

Sources: