Manus 2.0 launches with rebuilt agent framework and personal agent Cue

On September 28, the Manus team released Manus 2.0 and simultaneously released the standalone app Cue. According to official information, Manus 2.0 is already live on web, desktop and mobile; Cue has opened invitation-based early access on web, desktop and some mobile platforms, with the iOS app still awaiting App Store review.

Looking at this update as a whole, the problem Manus wants to solve already goes beyond "letting AI complete a single task according to instructions." An agent that can continuously handle things for a user needs to know when to start work, where to run, how to revise results together with people, and how to contact external services and other people.

So this release contains two layers of product claims: Manus Studio lets humans and AI complete and revise work together, while Cue lets AI participate in follow-up matters with its own identity. The former must prove that the results an agent delivers can withstand editing; the latter must prove that once an agent steps out of the chat box, it can still accurately understand authorization and reliably handle things on someone's behalf.

The underlying update of Manus 2.0 is called Cascade, a new iteration of its in-house agent framework. According to the official introduction, Cascade lets projects start from a lightweight state and introduce the corresponding specialized capabilities only when a task requires them; briefs, pages, videos and automation workflows produced within a single project can also stay linked, rather than each becoming an isolated task result.

Manus provided a set of test numbers: in one test configuration, compared with the previous system, Cascade reduced Token consumption by 23.2%, shortened task completion time by 28.2%, and lowered running costs by 32%. It must be made clear that these are results the company obtained under a specific configuration, and it cannot currently be directly inferred that all users and all tasks will achieve improvements of the same proportion.

But its design intent is clear: as tools and capabilities multiply, an agent cannot stuff the entire capability set into context every time it executes a task, otherwise response time and running costs will rise along with it.

The second underlying capability is cloud computers. In the past, an agent could generate a website or game for a user, but if the project needed a server that stays online continuously, the task would extend into hosting and operations. Manus now allows users to purchase a dedicated cloud runtime environment for projects that need one; official examples include the servers behind multiplayer online games and long-running automation workflows. In other words, after AI finishes making something, there is still a place where what it made can keep running.

The third update is Automations. Previously, scheduled tasks only began when a preset time arrived; now, events occurring in connected services can also start work. For example, receiving a new email, fluctuations in ad data, a new calendar event, or updates in Slack or Notion can all become trigger conditions. Users describe in natural language what to watch for and what to do once it happens, and Manus claims it can build and run the corresponding workflow accordingly.

Taken together, these three updates change the lifecycle of a task: an agent can invoke capabilities on demand, be awakened by external events, and keep executing continuously in a cloud environment. It no longer can only wait for the user to open the chat box and issue another command.

If Cascade and cloud computers determine how an agent works, Manus Studio deals with how humans take over its results. Manus upgrades the desktop app into a workspace shared by humans and AI, covering documents, spreadsheets, PDFs, slides, as well as websites, code, video and games. Among these, the newly launched video editor and game development environment are the most obvious creative-capability expansions in this update.

The video editor emphasizes "editable." In Manus's product description, AI can first produce a rough cut, but video clips, images, text, motion graphics and audio appear on the timeline as independent assets. Users can swap the soundtrack, replace images with their own, move animations, then export the finished video, and can also hand the revised project back to Manus so it continues working on top of the existing changes. Its target scenarios include 30-to-60-second product ads, animated data charts, tutorials, vlogs and motion-graphics-driven videos.

The editor also offers an Alchemy mode: users can first provide a creative direction, then Manus participates in creative design while invoking video generation and code generation capabilities, after which users can still return to the timeline for fine-tuning. Two things need to be distinguished here: Alchemy mode does invoke video generation capabilities; whereas Xiao Hong mentioned in a post after the release that the Cue-related video was made by having the agent write code and then converting it into video, without using a video generation model. The two describe different production paths and must not be conflated.

The game development environment, meanwhile, tries to push "generate a playable prototype" further into modification and publishing. Manus says this environment can work in combination with video, image and programming models; users can view the game running in the editing panel while managing assets, modifying code, and adjusting character movements, art and scenes. Once done, the game can be published as a website, or a continuously online server can be provided via cloud computers to invite friends into a multiplayer game. Officially listed supported scenarios include racing, combat, competitive and 3D games.

Manus Studio also has an update easily overshadowed by the video and game capabilities: remote control and Computer Use. According to official instructions, users can make requests to Manus on their phone, having it find files, export PDFs or test apps in a connected, authorized computer session, while watching its desktop operations. Manus says the agent operates in an independent and visible workspace, and uses only the files, browsers and apps the user approves. Two kinds of "computer" take shape here: one is the cloud computer used by agent projects, responsible for continuously online running tasks; the other is the user-authorized local computer workspace, responsible for handling files and apps that were already on the user's device.

The most debatable and also most imaginative part of Manus 2.0 is the simultaneously released standalone app Cue. According to the official introduction, Cue and Manus use the same infrastructure, but it is oriented toward personal agent scenarios. In Cue, every agent has its own email, phone number, wallet and computer. It can send messages, make payments within a budget set by the user, complete tasks on its own computer; it can also answer calls on the user's behalf and leave call summaries. Users can also put several agents into the same group chat: one looks for event venues, one organizes a shortlist, another drafts a presentation, and finally the human decides the direction and the result.

The most critical change here is "identity." Previously, agents mostly borrowed the user's accounts and completed an operation inside an app the user had opened; Cue wants each agent to become an object that can be contacted individually. It has a phone number, meaning calls can be made to it; it has an email, meaning work can be sent to it; it has a wallet and a computer, meaning it has execution resources within a certain range. These designs do not mean the agent has a person's legal status, but they make it look more and more, in product interactions, like a role that is entrusted by someone and can continuously participate in matters.

Cue is currently open via invitation codes, as a time-limited free experience. Some early-access users shared the registration process and pricing details on x: the Cue mobile app added a step to register a phone number for the Agent, one of which is a US phone number at $0.99 per month; the Linux version of the cloud computer starts at $39 per month. However, according to the latest user feedback, the $0.99/month tier has already been claimed out, and the price has reached $9.99/month. These pricing details come from early-access user shares, and whether the official pricing and tiers are stable remains unclear.

The free version's quota resets every seven days, and it appears that experiencing the full functionality still requires topping up. Some netizens infer that a phone number that cannot send text messages is less a mobile identity marker than a voice shell for the Agent; this also leaves an unanswered question: after the account is canceled, can this number set up for the Agent still be retained?

Three weeks before Cue's release, Meta launched Muse. Muse is also positioned as a personal agent: it can run in a dedicated cloud virtual machine, comes with a browser, can connect to the apps users commonly use, send emails, book trips, fill out forms or handle shopping tasks. After the user closes the app, it can still keep working, and Meta also wants it to remember the user's goals and preferences and proactively offer suggestions.

Both Cue and Muse try to cross the threshold from "answering" to "getting things done," and both believe an agent needs a cloud computer. But the currently public product descriptions show different focuses: Muse revolves around a personal proxy that continuously understands the user, emphasizing cross-app task handling, long-term goals and proactive suggestions; Cue emphasizes multiple agents each with their own identity, letting them participate in division of labor through email, phone and group chat. This is a comparison at the product-route level, not yet a judgment on completion rates or which experience is better.

When agents start making payments, sending emails and answering calls, both routes will also run into the same hard problem: exactly how much power has the user granted it? In Muse's launch materials, Meta explains that users can choose which apps to connect and how much permission to grant; sensitive operations such as sending emails and shopping require confirmation, and an operation log is provided. Cue, meanwhile, proposes payments within budget, independent communication identities and multi-agent group chats. The closer these features come to real transactions, the more users need to know which step is completed by AI on its own, which step requires human approval, and how to trace things after an error.

On almost the same day Manus released 2.0, Meta also announced the creation of an enterprise platform business, planning to offer Muse, business agents, the Muse API and coding tools to enterprises and developers. This shows Meta is also trying to take the capabilities of personal agents beyond a single consumer app and into more service workflows provided by enterprises.

At the time of the 2.0 release, Xiao Hong also posted on WeChat Moments that Manus is actively preparing a domestic version, and mentioned that an open overseas ecosystem makes it easier to build product experiences. He wrote: "Manus was born earlier than Codex and Claude Cowork, and Manus has always been a general agent." For personal agents, ecosystem is not an abstract concept. An AI that can write plans and make videos can mainly rely on its own tools to complete work; an AI that has to order food, make calls and pay on someone's behalf must truly connect with external merchants, communication and payment services. The boundaries of its capabilities are, to a large extent, determined by these connections.

Personal agents also face a problem easily masked by product demos: when it handles matters, it always has to encounter other people. Will a restaurant accept an agent ordering on someone's behalf? Are merchants willing to communicate with it about quotes? Can the person receiving an email or call clearly know that the other side is AI, who it is entrusted by, and what it can promise? Having an email and a phone means the agent can enter the existing communication networks; but being able to enter the network does not mean the surrounding services and other people are already prepared to collaborate with it. Whether a domestic version can stand depends not only on the agent's own capabilities, but also on which services it can connect to, how it obtains authorization, and how the transaction and communication links are joined up.

The new version's features have already been listed richly enough. What truly determines how far it can go next is a more plain question: when users hand over their computer, contact information and part of their payment authority to AI, can it get things done, and also let users always know how things got done? Cue is still in an early experience stage, and Muse is still expanding its products and services; who can complete more tasks will keep changing in the short term.