Hello fam 👋

Our creed here is simple: cover AI before it matters. This newsletter is about spotting the signals that help you invest smarter, build faster, and stay ahead of the curve.

This week’s deep dive is on Manus AI, a startup chasing the holy grail: a real “worker agent” that actually delivers global productivity gains. You probably saw the launch hype a few months back. Now we have a clearer picture of what Manus is, how it works, and where it might go.

Can it work at scale? Let’s find out.

Back in high school, I binged John Green’s Crash Course World History videos on YouTube. Partly to cram for exams, partly because they made history feel like a story rather than a list of dates.

One episode stuck: Coal, Steam, and The Industrial Revolution. John’s simple but powerful point was that the first steam engines weren’t revolutionary. Thomas Newcomen’s initial design was basically a giant water pump, used to clear mines of flooding.

For decades, that’s all it was good for. Steam engines only became revolutionary when James Watt reimagined them as a general-purpose tool. Once it could power looms, trains, and ships, the modern age began.

That history feels hard to ignore when thinking about AI.

For the past two years, we’ve been in the Newcomen phase: marveling at the engines themselves. GPT, Claude, Gemini. The next leap isn’t only about making those brains bigger and better. It’s also about giving them hands.

That’s why when I saw Manus launch in March 2025, it felt like history clicking into place.

On March 6th, 2025, Manus launched.

Not a demo. Not a copilot. A real general-purpose agent, invite-only but powerful enough that people whispered: this feels like the ChatGPT moment for agents.

The frenzy was instant. Invitation codes popped up on second-hand marketplaces for thousands of dollars. Manus’s Discord swelled past 180,000 members in days. It felt less like a product launch and more like a Yeezy sneaker drop, except the prize was a digital coworker.

Early users shared videos of it booking flights, compiling analyst-style stock reports, or spitting out fully playable games in a few hours. For the first time, a computer could take a vague, multi-step problem and return a finished product.

It wasn’t perfect. Far from it. Early testers complained of looping errors, server crashes, and a tendency to “get lazy” on longer tasks. Others poked under the hood and accused it of being little more than a wrapper around Anthropic’s Claude and Alibaba’s Qwen. The debate quickly split into two camps:

But everyone agreed: Manus changed the conversation. Before March, agents were still a promise. After March, they were real. And just like ChatGPT, it forced everyone else to accelerate their own roadmaps and reimagine their positioning in the new agent economy.

Once the invite-code frenzy cooled, the obvious question surfaced: Is this thing real, or just clever plumbing?

The answer is messy. Manus is both a leap and a layer.

To the user, it feels like magic. You say: “Analyze Tesla’s financials, build a dashboard, publish the link.” Manus does it—end-to-end. No hand-holding needed. It feels like handing work to a competent (if occasionally flaky) intern.

Under the hood, though, there’s no singular “Manus brain.” It’s a multi-agent system stitched together inside a controlled sandbox.

Each Manus session spins up a mini digital workspace with a few specialized sub-agents:

This loop of plan, act, verify, repeat runs until the task is done. At the end, it hands you a finished artifact: be it a working dashboard, a hosted app, a document you could actually send to your boss.

Manus didn’t invent any new foundational models. At launch, it used Anthropic’s Claude 3.5 Sonnet plus fine-tuned versions of Alibaba’s Qwen, stitched together with open-source scaffolding like Browser-Use.

What makes it feel different is the orchestration. Manus turns models into teammates

Manus doesn’t rely on one model. It routes tasks dynamically:

Manus juggles cost and performance like a law firm: associates for grunt work, partners for judgment calls. The router decides who does what, and swaps in better brains as they drop. GPT-5 was live within days of launch.

Most early “agents” were brittle. They chained together steps in JSON/ReAct frameworks like search → read → summarize → repeat. It worked until the task got messy. One wrong turn, and they stalled.

Manus was built differently. Its edge comes from how it trains, adapts, and executes:

Every Manus task spins up a clean Linux workspace with a browser, terminal, file system, etc, inside what the team calls Manus’s Computer.

What that looks like in practice:

Watching Manus work in its sandbox makes it feel less like a chatbot and more like a junior teammate.

“What if anyone can operate a supercomputer from their home?”

In July 2025, Manus leveled up by introducing Wide Research Architecture.

Instead of one agent slogging through a task linearly, Manus can now spin up hundreds of agent instances in parallel. Think of it as an army fanning out: some scrape, some parse PDFs, others run models, then hand everything back to a central “editor-in-chief” that stitches the results into one coherent output.

From the user’s side? Nothing changes. You just say, “Analyze 50,000 clinical trial reports,” and get a structured brief minutes later.

Why does this matter? Two reasons:

It’s more than a feature. This is new infrastructure for knowledge work.

As someone who’s used OpenAI’s Deep Research for serious analysis, I had the same question you might: Is Manus actually better?

Manus Wide Research is designed to scan broadly across domains. It’s like an “automated analyst desk”, surveying many papers, news items, projects, and data sources to surface relevant signals. The goal is coverage and breadth. Prompts are often open-ended and exploratory.

OpenAI’s Deep Research is built to dig deeply into a single question. It runs extended reasoning chains, persistent context, and iterative exploration to produce a long-form, expert-level synthesis. The goal is depth and rigor.

With all that we now know about how Manus actually works, the obvious question lingers: is it still just a wrapper?

In a sense, yes. Manus is multiple LLMs, a sandbox, and Browser-Use stitched together. Clever engineering, but not a new block of science.

But history suggests orchestration often is the breakthrough. The graphical user interface didn’t invent computing, but it made personal computers usable. TCP/IP wasn’t new physics, but it turned a tangle of networks into the internet. Apple didn’t invent the touchscreen, Tesla didn’t invent the battery. The leap was making the parts work together in a way people actually wanted.

I think that’s Manus’s move: extreme repackaging that feels like the first digital worker.

“Extreme repackaging”.

That’s the pattern you keep finding when you trace founder and CEO Xiao Hong’s work back.

Xiao, better known by his nickname “Red”, came out of China’s hyper-competitive tech scene, where winning wasn’t always about inventing something totally new but about moving fast, remixing what existed, and packaging it into something people actually used.

His first venture, Nightingale Technology, built WeChat plugins like Yi Ban and Wei Ban Assistant, small helpers doing repetitive office tasks. They weren’t glamorous (basically glorified WeChat macros), but they somehow ended up on >2 million office computers and drawing investment from Tencent and ZhenFund.

Xiao’s next act, Monica.im, took the same playbook global. Launched as a Chrome extension in 2022, it aggregated GPT-3 and later GPT-4 into a single assistant for drafting emails, summarizing articles, and answering questions. It became one of the most widely used AI browser plugins worldwide, proof that his packaging strategy worked outside China too.

His signature pattern is clear: choose existing tech, package it, make it feel like a product instead of a research tool.

Manus is the purest expression of Xiao’s philosophy. A decade of learning to repackage until something genuinely useful emerges.

In hindsight, Monica was the prototype.

It was Xiao’s first real attempt to package large language models into a usable product for everyday people.

But by late 2024, Xiao and his team had realized the ceiling. Monica was still an assistant: reactive, single-step, stuck inside Chrome. To build what Xiao really wanted which was a system that could plan, act, and deliver end-to-end results, they needed to go beyond extensions and build a general-purpose agent platform.

That's how the Monica team became the initial Manus team. Xiao focused on product and strategy, while co-founder Ji Yichao (ex-Peak Labs, developer of the Magi search engine) was the technical core. They brought the Monica DNA into a system that really does work end-to-end.

Monica was the appetizer. Manus is the main course.

Manus has always been about wrapping existing tech into something usable. But its business strategy was about unwrapping itself from China.

When Manus launched in March, it was technically a Beijing startup. By August, it had zero employees left in China.

The headquarters shifted to Singapore, with satellite offices in Tokyo and San Mateo. The founders are Chinese, but the company now positions itself as a global SaaS business, not a China-first AI firm.

Why the dramatic move?

China was untenable. Regulations banned Western models like Claude, forcing Manus to experiment with domestic alternatives that never took off. U.S. chip export controls blocked access to the GPUs needed to scale.

And the revenue math was brutal: Western users pay 5x times more than Chinese ones, and billing in USD instead of RMB multiplies margins. For a company burning through compute like venture capital, the choice was obvious. Singapore offered capital, compliance flexibility, and hardware access in one package.

Of course, it got messy. Chinese critics saw betrayal. U.S. regulators opened a review into Benchmark’s investment under outbound AI rules.

And Manus isn’t alone. I’ve noticed a quiet trickle of China startups unplugging from Beijing and reincorporating elsewhere. TikTok carved out TikTok Global to reassure U.S. regulators. Shein shifted much of its corporate structure to Singapore ahead of its IPO. Even in AI, several China-born infra and agent startups have quietly “flipped” their corporate homes to Singapore or the U.S. to access capital and chips.

The old dream of “build for China and the world” is dead.

The AI supply chain is bifurcating: U.S.-aligned vs. China-aligned. Founders have to pick. And if your product eats GPUs for breakfast, the choice is obvious.

This may be the next wave: China-born founders going global by leaving China behind. That may be good for founders. It’s bad news for China’s tech ambitions.

Singapore, Dubai, even Tokyo could become the neutral staging grounds for AI companies that want access to both capital and compute without the regulatory baggage. China, despite its massive market and talent pool, risks losing its most ambitious founders to friendlier jurisdictions

The real test for Manus is whether it holds up in messy, everyday workflows.

Benchmarks can usually be a distraction, but for agents there’s one that matters: GAIA, the “General AI Assistants” test suite designed to mimic real-world tasks. Its a mix of research, reasoning, and multi-step execution problems.

On GAIA, Manus beat OpenAI’s Deep Research across every difficulty level. Level 1 (basic tasks), Level 2 (intermediate), even Level 3 (complex, multi-step workflows). Manus scored higher across the board.

But benchmarks only tell half the story. In practice, Manus is strongest where the loop of plan → execute → verify shines:

In other words: Manus isn’t an omnipotent digital worker. But it’s already scary good for text, code and data.

I don’t like to judge products based on benchmarks. They are good to know but often gameable. What matters is whether people actually use the thing.

So far? They are.

Users have thrown Manus at everything: dashboards, research projects, hiring pipelines, slide decks, marketing campaigns. No single killer app, just a steady stream of “wait, it actually did that?” moments.

I tried it myself with a Reddit sentiment analysis project. With a single prompt, Manus spun up its sandbox, opened a CLI, and began writing code for the workflow. Midway through, it paused to ask for Reddit API access and when I didn’t know how, it walked me through the setup step by step.

I got stuck on Reddit’s end, but the process itself was illuminating: Manus was actually working, narrating progress, and asking for input like a junior teammate.

What struck me most was the visual experience. I could literally watch Manus work inside its own computer. It’s oddly satisfying, like watching a timelapse of an office intern doing three days of grunt work in three minutes.

Browser tabs opening, commands firing in the terminal, scripts being written and executed in real time.

And for a second, I caught myself wondering: is this what AGI feels like at the edges?

Because for the first time, it didn’t feel like I was prompting an LLM. It felt like I was watching an intelligent system think and act on its own.

And I’m far from the only one. A doctor tweeted about asking Manus to make slides explaining CAR-T cells. Minutes later, it produced a polished, seven-slide deck that could have come out of a biotech conference.

Another user showed off the new Wide Research feature by asking Manus to meta-analyze 100 scientific articles on AI in medicine, pulling titles, abstracts, authors, journals, citations, institutions, etc into a clean, structured table, sorted by impact.

What would take a grad student weeks came back in minutes, with a regional breakdown of research output that could slot straight into Notion, Excel, or a news article.

That’s the pattern: Manus compresses workflows that normally take teams, tools, and time into a single agent run.

The team has leaned into this by curating a set of playbooks, basically task templates that capture what early users are doing and package them for everyone else.

Want to build an AI-powered website? Spin up the Website Builder playbook. Need a quick competitive analysis? Use the Market Research tool. Other playbooks cover slide generation, AI video production, Chrome extension building, even sketch-to-photo conversion.

The playbooks are more than marketing collateral. They’re onboarding ramps. They turn Manus from an intimidating blank box into a menu of “things you can try right now.” And each successful run nudges users closer to integrating it into their daily workflows.

In other words: the hype wasn’t just screenshots. It’s sticky because people are using it to do real work.

One thing that stood out to me, after scrolling through demos and reviews, is just how B2C-focused Manus feels right now.

This isn’t Fortune 500s signing multi-year contracts. It’s consultants, indie hackers, doctors, analysts, small agencies. People with no procurement team to slow them down.

And it makes sense. Manus is priced and packaged like a consumer product. You don’t need an enterprise sales call or a six-month integration timeline. You just log in, spin up your sandbox, and watch it work. That’s why the screenshots, testimonials, and viral threads are coming from individuals, not institutions.

But that doesn’t mean enterprise teams won’t adopt it. Yes, many enterprises will lean toward specialized agents fine-tuned for their workflow. But plenty of team-level workflows don’t need custom-built AI, they just need a competent generalist.

So, “enterprise adoption” will probably not happen top-down. Rather, it’ll seep in bottom-up, as people discover Manus can make the boring parts of their jobs easier.

That quiet adoption may already be showing up in the numbers.

At a recent Stripe Singapore Tour event, the Manus team shared that they’ve hit a $90 million annual revenue run rate, just five months after launch, with $100M already in sight. For a company barely out of invite-only beta, this is insane velocity.

Where’s that revenue coming from? The team didn’t say, but we did some snooping. The geography tells a story.

Their dramatic exit from China and relocation to Singapore tells us that Manus is chasing Western and pan-Asian markets rather than trying to straddle both. A Benchmark investor hinted that Japan has been one of the fastest-growing footholds, and Manus opening a Tokyo office after leaving Beijing makes that bet explicit.

And while it’s not the cleanest signal, web traffic does offer a rough sketch. Over the last six months, Brazil actually led all inbound traffic, with the U.S. close behind at ~10%. China came in third, though much of that may be legacy usage from before Manus pulled the plug on its domestic presence.

In other words, the early adopters are scattered across emerging and developed markets alike. Manus is global by necessity, and by design.

Manus was first through the wall. But by August 2025, it’s not alone. The “brains in a box” era is fading. Everyone’s chasing the hands.

In my opinion, the biggest sin you can make in AI is thinking of it as a zero-sum game.

That’s a big no-no.

The potential market is too big for it to be so. Different agents are carving out different niches: OpenAI leaning enterprise, Manus chasing individuals and small teams, GenSpark trying to win on price/performance. Millions of users are testing workflows, and there’s plenty of room for multiple winners.

But Manus’s real competition may not be other agents..

Zoom out, and the true rival is people.

For years, the go-to solution for busywork was hiring a virtual assistant (VA), usually offshore, often through agencies or freelance platforms. Inbox triage, data entry, scheduling, lead gen, light research, all the repetitive but necessary tasks that clog up a founder’s day.

A good VA could cost anywhere from $1,500–$2,500 per month, a fraction of a U.S. salary but still a real line item for startups or solo operators.

Manus threatens to eat that entire category.

Yes, humans still have edges. Taste, creativity, open-ended judgment. But for structured digital work, the gap is closing fast.

And the market at stake here is no rounding error. The global VA and freelance assistant industry is worth an estimated $10–20 billion annually. Millions of remote workers across India, the Philippines, and Latin America rely on that income.

The math is brutal. A $15/hour VA vs. a $0.50/task agent? And Manus doesn’t ask for coffee breaks, national holidays, or Monday motivation playlists.

In mid-2025, Manus closed a $75M Series A led by Benchmark, at a $500M valuation, a 5x step-up from 2024. Benchmark joins Tencent and Sequoia China as early backers, though the center of gravity has clearly shifted West.

That Benchmark round also triggered scrutiny. U.S. officials are reviewing whether outbound investment into AI agents with Chinese founders fits the new Treasury rules. If the review goes south, Benchmark could be forced to unwind. For now, though, Manus has the cash and, just as importantly, the validation to scale globally.

Putting my operator hat on, if I were Manus, I would double down on a strategy I call “Desk-to-Boardroom”. Start where the slope is easiest: with individuals.

Enterprise AI sales are trench warfare. 6 - 18 months of procurement cycles, compliance reviews, and endless pilots that rarely see daylight. Agents, on the other hand, can prove their worth to a single person in under a minute. That asymmetry is everything.

History backs this up. Slack spread inside teams until 79% of the Fortune 100 were already using it by the time the first big contracts got signed. Dropbox followed the same curve: 90% of its business accounts started as personal ones. Notion, Airtable, Figma all did variations of the same strategy. Win individuals first, and the enterprise sale comes to you.

Manus has the same opening. Knowledge workers don’t want ten vendor-locked copilots chained to individual apps. They want one agent that moves with them—through Gmail, Slack, Notion, Chrome, wherever the work actually lives. If Manus can be that companion, it positions itself not as a tool but as the default personal work agent.

The growth loops are obvious. A lawyer builds a Manus workflow for summarizing case files and piping updates into Slack. A colleague copies it. Soon the entire team is running it. Shared workflows spread virally, compounding growth without paid spend.

Yes, there are risks. Individuals churn faster and pay less. ARPU for consumer productivity SaaS tends to be $5–10/month, compared to $30–100+/seat in enterprise.

Giants like Microsoft and Google will 100% attempt to crush the category by bundling agents into Office or Workspace. This is probably the biggest existential risk for Manus. The good news is that Microsoft Co-pilot isn’t any real good…yet. The moat will come down to interoperability and execution speed.

In short: if Manus can capture even 1% of knowledge workers (~5M) at $10/month, that’s $600M ARR before touching enterprise.

From the top down, the numbers are just as compelling. The AI agent market was about $5.4 billion in 2024, but forecasts have it jumping nearly tenfold to $50–52 billion by 2030, reflecting a ~45.8% CAGR between 2025 and 2030.

Even if growth were to slow to a more conservative ~35% CAGR after 2030, the market could still expand to roughly $150 billion by 2034, underscoring just how massive this category could become.

The corner Manus cares about—general-purpose agents—isn’t broken out in analyst reports yet. Industry forecasts today only cover the broader AI agent market, so these numbers are our modeled assumptions.

General-purpose agents are still a niche today, I assume ~10% of the overall market, about $0.5B. This is mainly because most of the spend is in vertical copilots, chatbots, etc. By 2030, as autonomy improves and adoption broadens, we assume they could grow to around $12B (~24% of the total market), increasingly competing with both specialized agents and parts of the VA/freelancer market. Beyond that, growth could balance out as enterprise and sector-specific agents scale too, leaving general-purpose tools at roughly $35B by 2034 (~25%), a stable quarter of the market.

For Manus, this creates a range of possible futures:

Now of course 2034 is very far away. These numbers are more to give you a sense of the scale involved rather than any exact financial modelling.

The real test is whether Manus can build the platform stickiness and community flywheel that turns a tool into the default. If it pulls that off, Manus could own the “work agent” market, the way Notion owns productivity software.

Manus will never match OpenAI or Anthropic on capital or distribution. Its edge has to be velocity. If it can keep shipping new features (Wide Research, video generation, or whatever comes next) faster than the giants, it can hold user attention even without monopoly reach.

History is full of these “fast mover beats the giant” stories. Figma outpaced Adobe by shipping multiplayer design, browser-native tools, and relentless quality-of-life updates while Adobe clung to its PDF empire. Today, Figma sits at a $25B valuation, about 20% of Adobe’s market cap.

For Manus, Speed is strategy.

And geography helps. If OpenAI dominates the U.S. and Europe, Manus can sprint ahead in markets like Japan, Brazil, and Southeast Asia, places where distribution isn’t locked, and where users are hungry for cheaper, faster tools.

General-purpose agents feel magical, but enterprises usually want reliability and narrow expertise. That’s where specialist agents win:

But generalists aren’t doomed by any means. Manus has the chance to play a different role: the orchestrator.

You give Manus a messy task like “prep me for a board meeting,” and it decides what’s needed. It might analyze numbers itself, call a specialist agent for code or legal review, then stitch everything together into a polished output.

If Manus can wedge itself into that role, then being a generalist won’t be a weakness. It’s the thing that lets it sit at the centre of workflows. Specialists do the deep work, but Manus stays in control as the hub.

And here, it has more freedom than the Big AI labs. OpenAI must route through GPT, Anthropic through Claude. Manus can be model-agnostic, plugging into whichever tool is best for the job.

I used to worry about the dependency risks for companies building on someone else’s models. What if OpenAI or Anthropic pulled access, walling off APIs to favor their own agents? But my fears have been largely blunted by the flood of strong, cheap open-source models, especially coming from China, like Qwen, Kimi, z.ai. In fact, these Chinese models are already “flippening” the US models.

Let’s conclude by going back to our earlier analogy (sorry if you’re not a history fan).

The steam engine only changed the world once Watt turned it from a pump into a general-purpose tool. AI feels the same. LLMs were the Newcomen pumps; Manus is the first glimpse of a Watt engine for digital work.

The race now is about trust and the entry point. Do agents come bundled from Big Tech, or from fast independents like Manus? What’s certain is that people want a colleague who doesn’t just chat but delivers finished work.

The open question is whether Manus itself becomes that Watt engine, or remains the Newcomen pump that only showed what was possible.

Either way, the age of general-purpose agents has begun.

Thanks for reading,

Teng Yan & 0xAce

If you found this interesting, please share it with a friend. That’s all we ask for.

Want More? Follow Teng Yan, Ace & Chain of Thought on X

Join our YouTube for more visual insights on AI

Subscribe for timely research and insights delivered to your inbox.