You type a single goal into a box — research a market, build the spreadsheet, write the report — and then you close the laptop. Hours later, a finished file lands in your inbox. That is the promise at the center of Manus, and this autumn the company pushed it further.
Most assistants still stop at advice. Manus is built to keep going: open a browser, run code, fill a form, and hand back a result a person could actually use.
The startup, founded in China and now operating from Singapore, says it has returned to independent operations and is betting everything on the agent (智能体).
What Manus actually does
Manus is not a model you chat with. It is a layer that takes a goal, breaks it into steps, and calls tools — a browser, a code editor, a file system — inside a cloud machine to get the work done.
The company describes its edge as careful task planning and long-context execution rather than a novel underlying model; it rides on top of existing large models (大模型) and adds the agent (智能体) layer that does the clicking and typing.
The workflow follows a simple shape: plan, execute, deliver. You set the task, the system works asynchronously while you do other things, and you get a notification when the output is ready. For research, finance, travel planning, and store analytics, that closed loop is the selling point.
The same design also invites overuse. Because the agent (智能体) works while you are away, it is easy to hand it sensitive tasks — a competitor analysis, a personal finance sweep — without watching the steps. The convenience is real, and so is the exposure.
Manus 2.0 and Cue
In late September 2026, the company announced Manus 2.0, described as its latest multipurpose agent (智能体), alongside a separate app called Cue. Cue is meant for people who want to build and deploy their own personal agents without writing the whole stack themselves.
The two moves point the same way. Manus wants to be both the finished agent (智能体) many people use directly and the toolkit others use to assemble their own. Bloomberg reported the launch as part of a wider push by the startup back into the agent market.
The comeback, in plain terms
Manus first drew attention in early 2025 as a general agent (智能体) that could demonstrably complete tasks, not just describe them. After a period of corporate upheaval, the company said it had returned to running itself independently.
That matters for trust. An autonomous agent (智能体) that touches your files and accounts is only as safe as the company behind it. A smaller, independent team has to win that trust the hard way, with clear data handling and a working product.
Why it matters globally
The agent (智能体) race is no longer only about who has the biggest model. It is about who can wrap a model in enough tooling that a normal person gets a result without a prompt-engineering course. Manus, born in China's startup scene and aimed at a global user base, is one of the clearest examples of that bet.
If agents (智能体) become the default way people use software, the winners will be the teams that make the boring chores disappear, wherever they are based.
Buy or build
Manus is one answer to a question every team now faces: buy an agent (智能体) service, or assemble one from a model plus open tools. The buy option trades control for speed; the build option trades time for ownership. Manus sits firmly in the buy camp, which is why its reliability and data terms matter more than its feature list.
The gap between demo and daily use
A polished launch video is not the same as an agent (智能体) you would trust with payroll. The hard part of this category is consistency: a task that works once in a demo can fail on the tenth run because a website moved a button or a form added a required field. That is why Manus, and every team in this race, lives or dies on quiet reliability work that never makes the keynote. For readers, the signal to watch is not the feature list but how often the agent recovers when something goes wrong — and whether it tells you clearly when it could not finish the job.
Honest limitations
Manus runs in the cloud, so it cannot touch files stored only on your own machine unless you connect them, and sensitive material leaves your environment. The company remains unavailable to users inside mainland China. Independent, standardized benchmarks of task success rates are still thin, so claims about reliability rest on the vendor. An agent (智能体) that executes code and opens browsers carries real safety risk if guardrails fail.
What readers can do now
- If you are outside mainland China, try Manus 2.0 on a low-stakes research task and check whether the delivered file actually saves you time.
- Explore Cue only after reading its data-handling terms, since the agent (智能体) will act on your connected accounts.
- Compare Manus with open agent (智能体) frameworks before trusting any single vendor with sensitive work.