Manus Shows AI Apps Are Down to Three Things

@Yunxiao Guo

Last night, Manus released its 2.0 version — essentially a super-sized combo platter.

Let me quickly run through what's in this bundle. On the technical side: a new agent architecture called Cascade that's both better and cheaper, plus cloud computers and automation. For work and creative scenarios, there's the new Manus Studio, which can edit videos, generate videos, build online multiplayer games, and remotely control computers to get work done. And Manus's own One More Thing: Cue.

I jumped in to try it right away, especially this Cue.

Since I'd already done my homework on the Personal Agent concept, the first thing I did in Cue was connect my various accounts. After linking Gmail, GitHub, Notion, Instagram, Oura, and so on, I added my Codex memory backup plus a few of my Jike posts (and I'll say it again under my real name: Jike, please release an official connector).

These more or less cover most of my active expressions and passive records on the internet — a fairly complete picture of my context. My first request was: make something, anything, that shows me how well you know me. A few minutes later I received a visual letter. Honestly, it was pretty bad — the writing style was simultaneously AI-generated and cloyingly sentimental, so I'll spare you all the details. The only pleasant surprise was that Cue had browsed through my street photography on Instagram and pulled in other photos of that same city as illustrations.

Of course, for a Personal Agent, the more important experience lies in real-life scenarios like travel, dining, and shopping. An agent probably doesn't need to describe your taste with perfect precision; it needs to execute well on buying groceries or ordering takeout. Clearly, Cue was designed with this philosophy in mind.

When you create an agent in Cue, it first assigns that agent a cloud computer, a payment wallet, an email address, and a phone number. Red Xiao himself said that something intelligent enough, once equipped with all of the above, "could perhaps be called a 'person.'"

In the year and a half since Manus 1.0 launched, agent-related concepts have gone through many iterations. But the Manus team seems to have remained fixated on building a general-purpose, human-like agent product. And Cue, as envisioned, really does resemble a person. When it can do what a real human butler or assistant can do, would you ever have a moment of doubt that there's an actual person sitting behind Cue?

Cue also reminds me of the discussions around cloud-based OpenClaw from earlier this year. The lack of local context was a major criticism at the time, but now everyone understands that context scattered across various apps is likely higher quality than local context. Assuming you can get that context, the cloud isn't a disadvantage. Another variable is Computer Use — a capability that's advanced by leaps and bounds in recent model versions. Agents no longer need to wait for an app to open a CLI or API; they can directly operate software on their own cloud computers.

I also appreciate those signature Manus touches — like the iOS QR code scanning feature I haven't even used yet, probably a small experiment in moving agents to mobile.

But all of this somehow still feels insufficient. Even setting aside the fact that I couldn't fully experience the ecosystem, Cue and Manus Studio in their current form aren't stunning enough. The likely reason: since Manus first launched, everyone has gradually grown accustomed to everything agent-related, and technological spectacle isn't so easily reproduced. On top of that, everyone holds Manus to sky-high expectations.

This happens to be a good place to quote an official line from this 2.0 launch: "Think back to your first 'wow' moment: an idea, turned into reality right before your eyes."

For Chinese AI application companies, Manus has been a game-changer.

After large language models, the Agent concept started trending, but the wave of startups emerged at scale only after Manus burst onto the scene in March 2025. The temperature has fluctuated hot and cold for a year, but most of these startups failed to advance further on growth. Looking back at this year and a half, the most underrated variable has clearly been the leap in model capabilities and the ambitions of model companies. The application startups that have actually broken through remain just Manus, Liblib, and Genspark — the so-called "big three." And what's been discussed most recently are actually Tipsy and Emochi, which suddenly emerged in just the past few months.

The broader startup scene fell quiet quickly. Some typical boom-then-bust patterns emerged: product-manager-type founders, A2A social networks, proactive agents, pivoting to FDE, and so on. Some paths remain inconclusive, like whether user data and post-training can take application startups anywhere meaningful.

A month ago, elsewhere produced an AI Application Midterm Exam, and we received nearly 1,100 responses. Respondents generally sensed a warming of AI application investment, but beyond continued funding for a few top companies, this still lacks evidence of funds putting real money on the line. Most attribute it to capital rotation and valuation resets rather than demand, retention, and revenue having actually worked out. And beyond all this, the hardest question to answer remains the direction of model evolution.

But by now, what's relatively certain is that applications have essentially narrowed down to just three things: Coding / office work, video / games, and Personal.

Consensus first formed around productivity scenarios, mainly Coding and daily office work, but this has already entered the stage of all-out war among tech giants. Entertainment covers content production and content consumption platforms centered on video and games. Content production, for various reasons, has drifted toward production scenarios and into the crosshairs of big company products. So edgy companionship apps, 4399-style mini-games, and real-time interactive content consumption platforms have come in waves over the past few months.

In the past two to three weeks, from Muse to Today AI to Manus Cue, this wave of Personal Agents hasn't even lasted a month. After several products launched, everyone's now studying agents for life scenarios again.

Manus's 2.0 release this time is a rare full-suite offering among application startups. Beyond the aforementioned Cue, Manus Studio also prominently features video processing and game-making capabilities — you can see echoes of other leading AI applications, and they roughly correspond to the consensus categories mentioned above. This also includes directions that smaller companies have attempted, like ChatCut, which elsewhere previously covered. This is personally one of my favorite companies.

Moreover, equally or even more important than the 2.0 product iteration is Manus's announcement that it will assemble a team to develop products for the domestic Chinese market. This is also the most significant signal of this release.

For most application companies still in the startup phase, they must place their limited bets on one spot. Manus entering 2.0 no longer seems to be playing by these rules.

Cover image: Marcellin Auzolle, Cinématographe Lumière, 1896, Bibliothèque nationale de France