Manus 2.0 — a major update to the universal AI agent Manus, released on September 28, 2026. It includes five new features: the proprietary Cascade agent architecture, the Manus Studio workspace, the dedicated Cloud Computer cloud machine, event-triggered automations that run from email, calendar, Slack, or Notion, and a separate Cue app for personal agents with their own email, phone number, and wallet (the Manus blog). According to the company, in one test configuration Cascade used 23.2% fewer tokens, completed tasks 28.2% faster, and cost 32% less than the previous system. For business users, what matters more than the product itself is the shift it represents: AI agents are moving from a “ask and get an answer” mode to continuous event-driven work, with their own resources and budget.
This material is for business owners and leaders in marketing, sales, and operations who are evaluating universal AI agents. We break down what’s new in Manus 2.0, how it differs from similar products, what limitations Russian companies face, and how to build a similar automation on your own systems. Detailed instructions for using Manus are not included.
AIrassvet editorial team. Prepared on October 4, 2026 based on Manus’s official blog and industry publications verified on that date. The editorial team did not test the service itself; performance figures are provided by the company and are identified as such. The article was published on the AIrassvet website, which implements AI agents for business; the services section appears before the conclusion. There are no affiliate links or compensation from Manus in this article.
Quick Take
- Cascade — the new agent architecture: a project starts out “lightweight,” and specialized capabilities are connected only as needed.
- Event-driven automations: a task is triggered by a new email, a change in ad metrics, a calendar event, a Slack message, or an update in Notion.
- Cloud Computer — a dedicated environment that runs 24/7, even when the user’s laptop is turned off; purchased separately, price not published.
- Cue — an app where each agent has email, a phone number, a wallet, and a computer; the agent can make payments within the budget set by the user.
- For Russia: according to vc.ru, the service does not officially operate in Russia; there is no information in the announcement about security certifications or data retention policy.
Contents
- What Manus is and what changed in version 2.0
- Cascade: the new agent architecture
- Event-driven automations and Cloud Computer
- Manus Studio: a workspace for people and agents
- Cue: agents with email, phone, and wallet
- Manus 2.0 and other always-on agents
- Limitations for Russian companies
- How to build similar automation on your own systems
- Frequently Asked Questions
- Primary sources and verification method
- How AIrassvet helps implement event-driven AI agents
- Bottom line: agents are moving into always-on mode
What Manus is and what changed in version 2.0
Manus is a universal AI agent that takes a task in natural language, then plans the steps itself, works with the browser, files, and code, and returns a finished result: a report, website, spreadsheet, presentation, or video.
Manus has a complicated corporate history. In late December 2025, Meta acquired the company, but in April 2026 the Chinese regulator blocked the deal, and Manus announced that it was returning to independent operations; at the same time, the company said that, as required by law, it would delete some user data created after December 29, 2025 (The Next Web). For enterprise users, this is a clear example of risk: data in a third-party service depends on its legal fate. We wrote about this in our AI news roundup for the week through August 16.
Version 2.0 changes three things: how the agent is built internally (Cascade), where it works (Studio and Cloud Computer), and when it starts working (event-driven automations).
Cascade: the new agent architecture
Cascade is Manus’s proprietary agent harness: a software layer that manages the model, tools, and project context. According to the company, Cascade keeps the project lightweight from the start and connects specialized capabilities only when they are needed. The brief, page, video, and automation remain parts of one connected project (the Manus blog).
Claimed impact in one test configuration, compared with the previous Manus system:
| Metric | Change according to Manus |
|---|---|
| Token usage | −23.2% |
| Task completion time | −28.2% |
| Execution cost | −32% |
Source: the Manus blog. The company did not disclose the test conditions, so these numbers should be seen as directional rather than guaranteed for your tasks.
Why this matters beyond Manus. The principle of “loading tools and knowledge on demand” is the main way to reduce the cost of agents: every extra tool description in context is paid for at every step. The same approach underlies skills on other platforms; more details — in the article about the AI agent harness.
Event-driven automations and Cloud Computer
Previously, Manus could perform tasks on a schedule. In version 2.0 a task can start from an event in a connected service: a new email, a change in ad metrics, a calendar event, a Slack message, or an update in Notion (the Manus blog).
To keep such automations running around the clock, Manus introduced Cloud Computer — a dedicated cloud environment that keeps running when the user’s computer is turned off. The company gives examples: a server for a multiplayer game and an automation that needs to run continuously. Cloud Computer is sold separately; the announcement does not list a price.
What this can look like in business processes:
| Event | What the agent does | Where to be careful |
|---|---|---|
| An inquiry email has arrived | It parses the email, drafts a reply, and creates a customer record | The customer’s personal data is sent to a third-party service |
| Ad campaign performance has dropped | It gathers data, looks for the cause, and suggests changes | An agent should not be given the authority to change bids and budgets without approval |
| Calendar event (customer meeting) | Prepares a customer brief and agenda | Agent access to the CRM and correspondence |
| An update in the task database | Updates a report or project page | Data accuracy in the final report |
We described similar scenarios in Russian systems in articles on automating leads from email and on a daily executive summary.
Manus Studio: a workspace for people and agents
The Manus desktop app has turned into Manus Studio — a shared workspace for humans and AI for documents, spreadsheets, PDFs, presentations, websites, code, games, and video (the Manus blog). It now includes two professional environments:
- Video Editor — for creating short 30–60 second videos: product ads, user-generated content, training videos, and motion graphics; clips, images, text, and audio are edited separately on the timeline;
- Game Dev — a game creation environment that combines models for video, images, and code; multiplayer mode requires Cloud Computer.
Also added is voice remote control from a phone and a computer control mode that works in a clearly authorized space: only approved files, the browser, and apps.
For marketing, this is first and foremost a tool for rapidly producing short videos and assets. How to build a process like this with your own tools is covered in the article on an AI content factory.
Cue: agents with email, phone, and wallet
Cue is a separate app for personal agents, where each agent gets its own email, phone number, wallet, and computer (the Manus blog). Agents can correspond, make calls, and pay within the budget set by the user, then send a summary (The Next Web). In a group chat, agents can hand off tasks to one another. Cue is in early access and free with an invite code.
This is the boldest and riskiest part of the release. An agent with a wallet and a phone is no longer an assistant, but an executor with authority. Neither the Manus announcement nor The Next Web’s write-up on Cue describes the agent action confirmation process or safeguards (the Manus blog, The Next Web).
For a company, this implies rules that should apply to any agent allowed to act:
- a spending limit per operation and per period, not just an overall budget;
- human approval for payments, agreements, and emails to external recipients above a threshold;
- a separate agent account with minimum necessary permissions;
- a full audit log and the ability to revoke access quickly.
More details in the articles on the risk budget for AI agents and on managing autonomy.
Manus 2.0 and other always-on agents
Manus is not the only one moving toward “persistent” agents. Almost simultaneously, OpenAI introduced dots — always-on agents in ChatGPT with their own cloud computer and browser, access to more than 4,000 apps, and support for recurring tasks (9to5Google).
| Capability | Manus 2.0 | OpenAI dots |
|---|---|---|
| Own cloud environment | Cloud Computer, purchased separately | Cloud computer and browser for each agent |
| Launch by events and schedule | Email, ads, calendar, Slack, Notion | Recurring tasks, continuity between conversations |
| Separate agent “personality” | Cue: email, phone, wallet | Access through ChatGPT, Slack, or Teams |
| Availability | Web, desktop, and mobile versions are live; Cue is in early access | Pro and Business Premium plans in supported regions |
Sources: the Manus blog, 9to5Google. Details about dots are in the review of all OpenAI DevDay 2026 announcements, the architecture of long-lived agents is covered in the article about asynchronous AI agents.
Overall takeaway: an agent becomes an employee with its own environment, schedule, and permissions. The value shifts from a “smart chat” to how reliably the agent is embedded in processes and how tightly its rights are restricted.
Limitations for Russian companies
- Access. According to vc.ru, the service officially does not operate in Russia (vc.ru). Workarounds for registration and payment create additional legal and operational risks for the company.
- Data. Transferring customer personal data to a foreign service is a cross-border transfer under Law No. 152-FZ; more details are in the article about 152-FZ and foreign AI.
- Transparency. The Manus 2.0 announcement does not provide information about security certificates, data retention policies, or pricing for the new components.
- Vendor resilience. The story involving Meta’s deal and the deletion of part of the data shows that a process built entirely on an external agent can stop for reasons outside the company’s control; see the AI service exit plan.
This section is not legal advice.
How to build similar automation in your own systems
The ideas in Manus 2.0 can be implemented in-house — with data kept within your own environment and integrated with Russian systems. A typical architecture:
- Triggers. Events from email, CRM (Bitrix24, amoCRM), 1C, messengers, and ad accounts flow into an automation platform — for example, n8n; for pricing and capabilities, see the article about n8n for business.
- Agent. A language model with a limited set of tools: data reading, draft preparation, field updates. Tools are connected as needed — as in Cascade.
- Persistent environment. A company server or virtual machine instead of Cloud Computer: the agent runs around the clock, and the data does not leave the organization.
- Control. Limits, human approval for external actions, logging, and quality metrics.
- Model. A cloud model accessible from Russia or a model deployed on your own servers, depending on data requirements.
A comparison of ready-made agent builders is in the material about AI agent builders.
Where to start: choose one event that repeats dozens of times a day and currently requires manual work, such as a new lead in email. Describe what the employee does in response, what data they use, and what counts as a good result. That will be the first automation.
Frequently asked questions
What’s new in Manus 2.0?
The Cascade agent architecture, Manus Studio workspace with a video editor and gaming environment, a dedicated Cloud Computer environment, event-driven automations, and the separate Cue app for personal agents with email, phone, and wallet.
What is Cascade in Manus?
It is Manus’s proprietary agent architecture. It starts a project with a minimal set of capabilities and adds specialized tools only as needed. According to the company, in one test configuration Cascade used 23.2% fewer tokens and was 32% cheaper than the previous system.
How much does Manus 2.0 cost?
With the release of version 2.0, the company did not publish new pricing. Cloud Computer is purchased separately; its price was not disclosed in the announcement. Cue is free during early access with an invitation code.
Can Manus be used in Russia?
According to vc.ru, the service officially does not operate in Russia. For companies, the requirements of 152-FZ for cross-border transfer of personal data are also important, so for workflows involving customer data, it is safer to build automation in your own systems.
How is Manus 2.0 different from OpenAI dots?
Both products focus on always-on agents with their own cloud environment. Manus emphasizes event-driven automations, content creation, and Cue agents with email, phone, and wallet; dots are built into ChatGPT and connect to more than 4,000 apps.
Is it safe to give an AI agent a wallet and permission to pay?
Only with strict controls: per-transaction and time limits, human approval above a threshold, a separate account, and an activity log. The Cue announcement does not describe approval procedures, so this feature should be used very cautiously for business payments.
Primary sources and verification method
Facts were verified on October 4, 2026, using the following sources:
- Introducing Manus 2.0 (Manus): release contents, Cascade, automations, Studio, Cloud Computer, Cue;
- The Next Web on Cue: agent budgets, availability;
- The Next Web on the break with Meta: the deal history and data deletion;
- 9to5Google: OpenAI dots for comparison;
- vc.ru: availability in Russia.
Method: the product description was taken from the official Manus blog and industry publications; the service was not tested. Cascade metrics are company data without disclosed test conditions. Architecture recommendations are editorial analysis.
How AIrassvet helps implement event-driven AI agents
Manus 2.0 shows where AI agents are headed, but for company workflows, integration with its systems, data control, and limiting the agent’s authority are critical. AIrassvet develops AI agents integrated with communication channels and business systems, and automates actions in ERP, 1C, and browser interfaces. For the tasks in this article, four areas of work are especially useful:
- process audit: which events in email, CRM, and messengers repeat most often and which of them can be handed off to an agent;
- development of an MVP agent that starts on an event, works with CRM, 1C, or email, and passes actions above a set threshold to a human;
- agent-based automation of actions in software interfaces when the system does not have a convenient API;
- deployment of the model within the company’s environment so customer data does not go to an external service.
The first step is to choose one repeating event, describe the current process, data sources, constraints, and the criterion by which the automation will be considered successful. Discuss the task.
Bottom line: agents are moving into always-on mode
Manus 2.0 is one of the most revealing releases of fall 2026. The Cascade architecture shows how to make agents cheaper by connecting tools on demand; automations and Cloud Computer show how an agent becomes a continuously running executor; Cue shows how far the market is willing to go in handing agents authority, including payments.
For Russian companies, using the service itself in business processes is difficult: it does not officially operate in Russia, and the announcement does not disclose the data retention policy. But the ideas behind the release can be transferred to your own systems: an agent that starts from an event, runs within the company’s environment, has limited rights, and passes anything above the threshold to a human. You should start with one repeating event and a measurable success criterion.