How Manus Works Under the Hood

All plans2 min read

Step 1: Receive Goal

You provide a high level task description. Manus's orchestration agent analyzes the goal and determines what needs to be done.

Step 2: Decompose into Sub Tasks

The planner breaks the goal into a structured sequence of sub tasks with dependencies, priorities, and success criteria.

Step 3: Assign to Specialized Agents

Each sub task is routed to the best suited agent: browser agent for web navigation, code agent for programming, writing agent for content, retrieval agent for search.

Step 4: Execute in Virtual Environment

Agents work in a cloud VM with a real browser, code sandbox, and file system. You can watch progress in real time through Manus's Computer.

Step 5: Deliver Results

Finished outputs (files, reports, websites, code) are saved to your Manus Library for download. You receive a notification when the task is complete.

Specialized sub agents

🌐
Browser Agent

Navigates websites using a real Chromium browser. Handles search, form filling, data extraction, and multi page workflows.

💻
Code Agent

Writes, executes, and debugs code in a sandboxed environment. Supports Python, JavaScript, and full stack web development.

📝
Writing Agent

Generates prose, reports, documentation, and presentations. Handles formatting, structure, and style for professional outputs.

Multi model routing

Manus does not rely on a single AI model. It routes tasks to different LLM backends (including Anthropic Claude and Alibaba Qwen) based on which model performs best for each sub task type. This is why Manus can sometimes outperform any individual model on complex, multi step workflows.