Manus 2.0: A Comprehensive Overview of the New Agent Architecture

Introduction

Released on September 28, 2026, Manus 2.0 represents a fundamental architectural shift rather than a simple version update. The platform has evolved from a task-completion tool into a comprehensive execution environment where AI agents can maintain projects, operate persistent environments, and collaborate with human users. This release introduces a new agent harness, professional creation tools, persistent cloud infrastructure, and a new personal-agent product called Cue.

Benefits

At the heart of Manus 2.0 is Cascade, the new in-house agent harness. Unlike previous iterations that loaded every capability into every project, Cascade keeps projects lightweight by bringing in specialized tools only when needed. This orchestration approach has demonstrated significant efficiency gains in tested configurations with 23.2% fewer tokens used, 28.2% faster task completion, and 32% lower operational costs. Cascade ensures that a project remains connected from the initial brief through to the final deliverable, whether that involves documents, code, video, or automation.

Manus Studio is the upgraded desktop environment that serves as a shared workspace for people and AI. It integrates professional tools for various creation types, including documents, spreadsheets, PDFs, slides, websites, code, games, and video. The key innovation is continuity where the agent remains involved in the project throughout the creation process rather than handing off artifacts to separate applications. The platform also features a professional video editing environment with Alchemy Mode, an AI-driven creative director feature that uses video and code generation together to produce high-quality initial cuts. Users can build, test, and publish playable games with support for multiplayer functionality through integrated server hosting.

For projects that outgrow temporary sandboxes, Manus 2.0 introduces Cloud Computer, a dedicated, persistent virtual machine environment. This allows continuous operation of applications, bots, and scripts along with persistence of files, installed tools, and running processes between sessions. The automation capabilities have expanded from simple scheduled tasks to event-triggered workflows where agents can initiate work based on real-time events from connected services including new emails, changes in ad performance, calendar events, Slack messages, and Notion updates.

Manus 2.0 bridges the gap between the agent and the user's physical computer through Computer Use, which allows Manus to operate within an authorized, visible workspace on the user's computer using approved files, browsers, and applications. Remote Control enables users to view their desktop on a mobile device and issue voice commands to Manus while it works.

Cue is a new standalone application built on the same infrastructure as Manus, designed for personal agents with distinct identities. Available on web, desktop, and mobile, Cue allows users to create agents that function as ongoing operators rather than one-off assistants. In Cue, each agent possesses its own email address for communication, phone number for receiving calls, wallet for budgeted payments and transactions, and computer for independent task execution. Cue supports multi-agent teams where multiple agents collaborate in a group chat around a shared goal. The user sets the direction and makes final decisions.

Use Cases

Manus 2.0 is excellent for solo research involving open-ended web research, data gathering, and report generation. The platform is capable of planning and executing multi-step tasks without hand-holding and has strong capabilities in navigating websites and extracting data. The video editing environment is ideal for creating short product ads, AI UGC, tutorials, and motion graphics. The game development tools enable users to build, test, and publish playable games with multiplayer functionality. Cloud Computer is perfect for hosting multiplayer games and long-running automations that require continuous operation. Cue is designed for everyday services integration where an agent can scan a QR code at a restaurant to order food or hold a place in line. Multi-agent teams can be used for complex projects where one agent researches venues for an event, another creates a shortlist, and a third drafts the presentation.

Pricing

Manus 2.0 is available now on web, desktop, and mobile. Manus offers Free, Pro at $20 per month, and Team plans at $20 per seat. Cue is in early access and is free to use with an invite code. Cloud Computer is available as a paid add-on with pricing at $10, $30, or $50 per month.

Vibes

Manus has undergone significant corporate changes in 2025-2026. The platform had a viral launch in March 2025 with an invite-only waitlist that exceeded 500,000 users. In December 2025, Manus was acquired by Meta for approximately $2 billion. However, in April 2026, China's National Development and Reform Commission ordered the deal unwound due to foreign investment restrictions. In August 2026, Meta cut off access to its data and Manus announced it would return to independent operations. On September 1, 2026, Manus resumed independent operations under new investors including Tencent, ZhenFund, and HSG who bought the company back for $2 billion. Some users experienced data deletion between August 23-25, 2026 as part of the separation from Meta, but a restore portal was opened.

Additional Information

Manus 2.0 represents a maturation of the AI agent concept, moving from simple chatbots to persistent, capable workers. The introduction of Cascade, Cloud Computer, and Cue provides a robust foundation for both professional creation and personal automation. While the platform has faced corporate turbulence, its return to independence and the release of this comprehensive architecture position it as a significant player in the agentic AI landscape. Users should consider their specific needs for persistence, collaboration, and autonomy when evaluating Manus 2.0 against alternatives. The platform has some limitations including memory management where tasks do not inherently write learned information back into a shared memory pool across sessions. Team features exist but may not offer the same depth of shared workspace memory as dedicated multi-agent platforms. Complex multi-step tasks with 8 or more steps can face reliability issues common to autonomous agents. Variable task costs also make budgeting challenging.