Manus AI: The Chinese Agent's 2026 Saga & Lessons for US Businesses

Manus AI returned to the top of search trends in the US for an unusual reason: the Chinese autonomous agent took its own platform offline for about two days to delete user data and finalize its split from Meta. If you follow AI agents — or are evaluating one for your operations — this story offers practical lessons that go far beyond the corporate dispute between Silicon Valley and Beijing.

TL;DR

• Manus AI is the general-purpose agent from Butterfly Effect: you describe the objective in natural language, and it plans, executes in the cloud, and delivers the completed work.
• Meta announced its acquisition in December 2025 for approximately $2 billion; China's NDRC blocked the deal on April 27, 2026, and Meta terminated the transaction on June 15.
• On August 11, 2026, the company announced its return to independent operation. Between August 23 and 24, affected accounts were blocked for data deletion; restoration opened on August 25.
• Reuters reported in July that Tencent was negotiating to become the largest — though still minority — shareholder, at the same $2 billion valuation.
• For businesses operating in the US, the case serves as a manual on vendor lock-in, data portability, and privacy regulations like CCPA or state privacy laws for autonomous agents.

What is Manus AI and Why It's Back in the News

Manus debuted on March 6, 2025, in an invitation-only beta, and became a phenomenon due to a promise different from chatbots: instead of answering, it executes. You type "create a spreadsheet comparing 40 vendors and deliver a report," and the agent breaks the task into steps, opens a browser in a cloud environment, researches, writes code, generates the file, and returns the result. This is what literature calls a general-purpose agent — the paper From Mind to Machine: The Rise of Manus AI as a Fully Autonomous Digital Agent describes the system as a combination of LLM reasoning with end-to-end execution that produces tangible deliverables.

Behind the product is Butterfly Effect, founded by Xiao Hong with co-founders Ji Yichao and Zhang Tao. The company originated between Beijing and Wuhan and moved its headquarters to Singapore in mid-2025 — about 40 of its 120 employees in China accompanied the move. In April 2025, the startup raised approximately $75 million in a Series B, at a valuation of about $500 million. By December 2025, its annualized revenue was already around $125 million.

What brought the name back to Google Trends in the US in August 2026 was not a new launch. It was a corporate restructuring with a visible side effect for the end-user: days of unavailability and data deletion.

The Manus AI Timeline: Meta, Beijing, and the Unwound Acquisition

It's worth reading the timeline carefully, because each step has a parallel in any software contract you sign:

1. December 2025 — Meta announces the acquisition of Butterfly Effect for a reported $2 billion (some sources mention up to $3 billion).
2. January 2026 — China's Ministry of Commerce publicly reminds that companies in the country must comply with Chinese law.
3. April 27, 2026 — The National Development and Reform Commission (NDRC) prohibits foreign investment in the startup, as part of a broader move by Beijing to maintain domestic control over its AI companies.
4. June 15, 2026 — Meta officially terminates the transaction rather than fight the Chinese regulator.
5. July 10, 2026 — Reuters and Financial Times report that Tencent is negotiating to acquire the largest stake in the company, alongside ZhenFund, HSG, and the management itself, maintaining the $2 billion valuation. Even as the largest shareholder, Tencent would remain a minority owner.
6. August 11, 2026 — The company publishes a notice to users: "will return to operating as an independent company."

Amidst all this, in March 2026, the product also gained a desktop app to run the agent directly on the user's computer. In other words: the roadmap continued to advance while ownership control was disputed across two continents.

Data Deletion from August 23-25: What Actually Happened

This is the part that matters to anyone using Manus AI for work. To comply with "regulatory requirements in certain jurisdictions," the separation from Meta required permanently deleting data generated by some users since December 29, 2025. The published schedule, in Singapore Time (SGT), was as follows:

Date (SGT)	What Happened
August 11-23	Affected users were no longer charged
Starting August 23	Backup tool released for data download
Aug 23, 8 AM → Aug 25, 7:59 AM	Accounts and platform inaccessible (~2 days)
August 23-24	Permanent data deletion window
Starting August 25	Restoration portal for downloaded data

In practical terms: anyone with workflows supported by the agent needed to export everything within a short window, was without access for two days, and only recovered their history if they had performed a backup. The company mentioned "millions of users globally" — this was not a niche event.

And here's the first practical lesson, which I repeat in every architecture meeting: backup and export are not premium features, they are fundamental requirements. If your vendor doesn't provide a continuous export path, you don't own the data; you merely have a license to use it.

How Manus AI Works Under the Hood

Manus AI operates on a multi-agent architecture: one component plans, others execute specialized sub-tasks (navigation, code, information retrieval), and an orchestrator stitches the results together. It's the same principle that Google adopted in Gemini Spark and that Atlassian brought into Jira with its Rovo agents.

At launch, the startup claimed state-of-the-art performance in GAIA, the benchmark that measures agents in real-world tasks using tools, surpassing OpenAI's Deep Research across all three difficulty levels. By 2026, the landscape is much tighter: competing agents publish better numbers on the same benchmark, and leadership changes hands every quarter.

Why Benchmarks Shouldn't Drive Your Choice

I've learned from experience that benchmark rankings are a poor purchasing criterion. GAIA measures standardized tasks with generic tools. Your operation has internal systems, spreadsheets with odd names, undocumented business rules, and a WhatsApp with 3,000 open conversations. The agent that wins the benchmark doesn't necessarily win in your specific process.

What typically decides, in order: integration with your existing stack, cost predictability, data control, and the ability to audit what the agent has done.

Manus AI vs. Gemini Spark and Western Agents

There's no "best agent" — only the right agent for your risk profile and tech stack. Here's the breakdown I use to guide clients:

Criterion	Manus	Gemini Spark	Built-in Agents (Rovo, Copilot)
Origin / Jurisdiction	Butterfly Effect (Singapore, Chinese capital)	Google (USA)	Your existing SaaS vendor
Usage Model	Open-ended tasks, cloud execution	Continuous assistant with calendar and context	Agent within product workflow
Best For	Research, reports, one-off deliverables	Personal routines and productivity	Automating existing processes
Key Consideration	Data governance and ownership instability	Google ecosystem dependency	Limited to product scope
Cost	Credits per task	Subscription	Built-in/per-user license

The Manus AI credit model deserves its own paragraph. Each action consumes credits, and consumption scales with complexity: a short conversation uses few, while deep research with dozens of sources can consume hundreds of credits in a single execution. This is good for starting out (you pay for what you use) but challenging for budgeting (monthly cost becomes a function of team behavior). Before standardizing the tool across your team, run a two-week pilot measuring consumption by task type.

What the Manus AI Saga Teaches About Relying on an AI Agent

Five lessons I would take from this case, all applicable to any AI vendor — including American ones:

• Regulatory risk is product risk. An NDRC decision in Beijing led to two days of unavailability for a user in the US. Geopolitics isn't an abstract topic when your process depends on a foreign API.
• Agent-generated data is also your data. Reports, spreadsheets, and code produced within the platform need to have a copy outside of it. Export routinely, not in a panic.
• Critical workflows don't reside in a single tool. If customer service, billing, or report generation stop when a SaaS goes down, your design is flawed — a documented manual fallback is missing.
• Contracts need a portability clause. Export format, notice period, and data destination in case of termination. Without this, you're signing a blank check.
• Novelty is not strategy. Adopting the trendiest agent without mapping your process only accelerates existing chaos.

Where AI Agents Truly Deliver Value for a US SMB

Tools like Manus AI shine for one-off deliverables, but recurring returns come from elsewhere. At Agathas Web, the most common pattern isn't the generic agent solving everything — it's the narrow agent effectively solving an expensive step. Three areas where I see consistent returns:

WhatsApp Customer Service. Triage, qualification, and first-line responses with human escalation. Here, the infrastructure decision weighs more than the AI model: running on the Official WhatsApp API instead of the Business app is what ensures history, multiple agents, and lower risk of blocking.

Education and Online Learning (EAD). Generating content drafts, assisted correction, and summarizing forums within Moodle. The agent doesn't replace the tutor; it shortens the time between a student's question and a useful answer.

Marketing and Paid Traffic. Creative analysis, search term grouping, and drafting ad variations. This is where gains appear fastest because the test cycle is short and measurable.

In all cases, the rule is the same one I've applied for over fifteen years in software projects: automate the process you can already describe. If no one on the team can explain the rule in two sentences, the problem isn't a lack of AI.

Checklist: Adopting Manus AI Without Creating Operational Debt

If you're testing Manus (or any autonomous agent) this week, follow this order:

1. Choose a tedious, measurable task. Weekly reports, competitor research, consolidating spreadsheets. No "let's automate the entire customer service."
2. Define the pilot's cost ceiling. Credits or subscription, with a fixed monthly limit and someone responsible for monitoring consumption.
3. Never feed sensitive personal data into the pilot. Social Security numbers, medical records, banking details, and customer databases are off-limits until a contract and risk assessment are in place.
4. Export everything weekly. Automate the backup of deliverables to your own drive. The August case shows why this isn't paranoia.
5. Compare with the human baseline. Time spent before, time spent after, rework generated. Without numbers, the discussion becomes opinion.
6. Document what the agent did. Prompt, sources, and output. Auditing isn't bureaucracy — it's what saves you when the result goes wrong.

CCPA and Data Sovereignty: The Blind Spot for Autonomous Agents

An autonomous agent like Manus AI is, by definition, a system that sends your context externally and acts on your behalf. Under regulations like CCPA or other US state privacy laws, this places you in the position of data controller for the data you input into the prompt — and the foreign platform becomes the data processor, with international data transfers involved.

Three questions that resolve 90% of the problem before it becomes an incident:

• What personal data is in the prompt? If the answer is "I don't know," the real answer is "probably yes."
• Where is the data processed and for how long is it retained? A clear retention policy, with timeframes and methods for deletion.
• Is there a legal basis and record-keeping? Poorly documented legitimate interest won't withstand regulatory scrutiny.

It's also worth following the US regulatory movement on AI, which is moving towards requiring transparency and risk assessment in automated systems. Those who already organize contracts, data inventories, and usage logs now will spend much less later. For the context of corporate vendors and agents, it's also worth reading what I discussed about what Gemini Spark changes for businesses and about the Google I/O 2026 updates for US businesses.

Conclusion: The Tool Belongs to the Vendor, the Process is Yours

The story of Manus AI in 2026 is interesting for its plot — a billion-dollar acquisition blocked by a regulator, Tencent entering through the back door, two days of scheduled blackout — but it's useful as a reminder: no AI agent is stable infrastructure. They are fast, powerful, and replaceable layers over processes that need to be yours.

Test Manus AI. Test Gemini Spark. Test whatever appears next month. Just don't build the core of your business within an account that could become inaccessible for 48 hours due to a corporate decision on the other side of the world.

If you're evaluating where AI agents make sense in your operation — customer service, online learning (EAD), or paid traffic — it's worth starting by mapping just one process, with defined costs and metrics. It's the kind of conversation we have here every month, and almost always the first step costs less than you imagine.