Manus 2.0: From a general-purpose agent to creation and personal agents

Manus announced version 2.0 on September 28. The company describes it as more than a version update: a rebuild spanning a new architecture, products, and capabilities.[1]

This is not an independent product test. I have not compared Manus 2.0 with the previous version on the same task set. The feature and performance descriptions below come from Manus's announcement. My question is narrower: what changes for a user when a general-purpose agent expands into professional creation and a personal-agent layer?

The real change is outside the chat box

The first agent products were built around a simple expectation: give an instruction and wait for the result. Manus 2.0 proposes something more persistent. A project can have its own working environment, editable production tools, triggers, and, when needed, several agents.

Cascade is the foundation. Manus says its in-house agent harness keeps a project light at the start and loads specialised capabilities only when the work calls for them. The brief, page, video, and automation remain parts of one connected project.[1]

For the user, that means treating each output less like the end of a separate conversation and more like a project that can be reopened and revised. The agent starts to look less like a tool that answers and more like a working layer that carries context through the job.

Creation no longer ends with the exported file

Manus positions Studio as a shared workspace for people and AI. It covers documents, spreadsheets, PDFs, and slides, as well as websites, code, games, and video. According to Manus, Studio does not load every tool into every task; it brings in only what the project needs.[1]

The important part is that the work remains editable. Video Editor exposes clips, images, text, motion graphics, and audio as separate pieces on a timeline. A user can replace music, add their own footage, and then hand the revised work back to Manus. Game Dev similarly keeps the running game, assets, and code in one editing environment.[1]

That could reduce the pressure to write one perfect prompt. The agent makes a first pass; the user keeps the final decisions and detailed corrections. It is not a quality guarantee. I have not independently tested the examples in the announcement, so I cannot judge Studio's style consistency, resilience on large projects, or the amount of manual repair it needs.

Cloud Computer and Automations give work a longer life

Cloud Computer provides a dedicated environment for projects that need to outlive one session, such as a server or a permanent automation. Automations goes beyond time-based schedules: an email, calendar event, Slack message, Notion update, or change in ad performance in a connected service can start a task.[1]

Together, they move the command to “run the agent” into the background. Work can become a process that waits for an event, runs in its own environment, and produces a result. That creates a new set of questions: which service has which permission, where does the process stop after a failure, and who approves a step that sends a message or spends money?

The useful part of Manus 2.0 is not simply more autonomy. It is the chance to make boundaries explicit. Before automating anything, define the trigger, spending limit, human approval, and rollback path. The distinction in AI automation is not transformation matters here: speeding up a task does not fix a poorly designed process.

Cue turns the personal agent into a separate product

Cue is a standalone personal-agent app built on the same infrastructure as Manus. The company says each agent can have its own email, phone number, wallet, and computer; it can pay within a budget and answer calls. Several agents in one group chat can hand work to each other.[1]

This is the clearest step from a general-purpose agent to a personal-agent layer. Once identity, communication channels, a budget, and a computer come together, the agent is no longer just an interface that produces text. It can act as a representative in the outside world. The risk grows with the utility. A bad email, payment, or booking costs more than a bad paragraph.

Cue is in early access and free with an invite code. Manus says version 2.0 is available on web, desktop, and mobile, while Cue on iOS is awaiting App Store review.[1] Independent evidence is still needed on reliability, permission controls, and recovery from failed actions at broader scale.

A decision frame for users

“Is it smarter?” is not enough when evaluating Manus 2.0. A more useful sequence is:

1. Can the result be edited in detail after generation?
2. Does project context survive across different production stages?
3. Are the automation trigger, stop condition, and approval point explicit?
4. Can the agent's identity, budget, and service access be limited separately?
5. Can a bad action be reversed and audited?

The choice between Cloud Computer and a local machine also depends on the task. Always-on services fit the cloud; sensitive files, live debugging, or work needing frequent human intervention may belong locally. The local versus cloud agent guide goes deeper into that split.

Singrey's note

What caught my attention in Manus 2.0 was not the arrival of one more agent. It was the attempt to put the production tool, working environment, automation, and personal representative inside the same architecture. If that combination works, the skill to learn will be designing work boundaries rather than writing prompts. I am cautious for now: the direction is interesting, but company examples are not a substitute for independent testing.