Spotlight on Manus

Key Concepts

• Manas: An AI agent designed to perform tasks by interacting with a virtual computer, including a file system, terminal, VS Code, and a real Chromium browser.
• Cloud Models: Large language models (LLMs) hosted in the cloud, specifically Anthropic's Claude, used as the "brain" for Manas.
• Agentic Workflow: A process where an AI agent autonomously plans and executes a series of actions to achieve a goal.
• Tool Use/Function Calling: The ability of an AI agent to utilize external tools or functions to accomplish tasks.
• Less Structure, More Intelligence: The core philosophy behind Manas, emphasizing a simple, robust structure that allows the foundational model to improvise and handle complexity.
• Long Horizon Planning: The ability of a model to plan and execute tasks that require multiple steps and iterations.
• Coot Injection: A technique used to improve function calling performance by injecting reasoning from a planner agent into the main agent.

What is Manas?

Manas is an AI agent developed by Manas AI, designed to act as a "hand" for large language models (LLMs), enabling them to interact with the real world. The name "Manas" is derived from MIT's motto, "mens et manus," meaning "mind and hand." Manas aims to bridge the gap between the intelligence of AI models and their ability to perform practical tasks.

• Core Functionality: Manas provides a fully functional virtual machine to each user, including a file system, terminal, VS Code, and a real Chromium browser (not headless). This allows Manas to perform a wide range of tasks, such as unzipping files, extracting data from PDFs, browsing the web, and interacting with web applications.
• Use Cases: Manas can be used for various tasks, including finding office spaces, analyzing room styles and suggesting furniture from IKEA, and conducting deep research.
• Internal Use Case Example: Manas was used to find a suitable office space in Tokyo for 40 employees, including accommodation options. It provided an interactive map with 10 options, including office locations and nearby accommodations, along with prices and reasons for choosing each option. This was accomplished in under 20 minutes.
• User Use Case Example: A user can send an image of an empty room to Manas, which will analyze the room's style, browse the IKEA website for furniture, and generate an image of the room furnished with IKEA products, along with a document containing links to purchase the furniture.

How Manas Was Built

The development of Manas was inspired by the code editor Cursor. The founders observed non-coders using Cursor to solve daily tasks by simply accepting code suggestions without understanding the code itself. This led to the idea of building the "right panel" of Cursor, an AI assistant that runs in the cloud.

• Key Components:
  • Virtual Computer: Each Manas task is assigned a fully functional virtual machine with a file system, terminal, VS Code, and a real Chromium browser.
  • Pre-integrated Private Databases and APIs: Manas pre-integrates private databases and APIs, allowing users to access real-time financial data and other information without needing to write code or understand API calls.
  • Personal Logic System: Users can teach Manas how to solve problems and customize its behavior. For example, users can instruct Manas to confirm details before conducting research.
• Development Timeline: The initial version of Manas was built in five months, from October to March.

Less Structure, More Intelligence

The core philosophy behind Manas is "less structure, more intelligence." This means that Manas has a simple, robust structure with zero predefined workflows, allowing the foundational model (Claude) to improvise and handle complexity.

• Emphasis on Context: Manas focuses on providing more context to the model rather than imposing strict controls or predefined workflows.
• Rejection of Multi-Role Agent Systems: Manas avoids using multi-role agent systems, where different agents are assigned specific roles (e.g., coding agent, search agent), as this is seen as limiting the potential of LLMs.
• Improvisation by the Model: Manas allows the model to improvise and adapt to different tasks, leading to emergent capabilities.

Why Anthropic's Claude?

Manas chose Anthropic's Claude model for three main reasons:

1. Long Horizon Planning: Claude is better at planning and executing tasks that require multiple steps and iterations compared to other models, which tend to answer questions in a single turn.
2. Tool Use: Claude has excellent tool use capabilities, which are essential for Manas to interact with the virtual machine and perform tasks. Manas initially used a technique called "coot injection" to improve function calling performance, which was later supported natively in Claude 4.
3. Alignment with Agentic Usage: Anthropic has invested significant resources in aligning Claude with computer and browser usage, making it well-suited for building agents.

• Token Usage: Manas spent $1 million on Claude models in the first 14 days of operation.

Browser Interaction Details

When Manas uses the web browser, it sends three pieces of information to the foundational model:

1. Text in the Viewport: The text content of the current browser viewport.
2. Screenshot: A screenshot of the current browser viewport.
3. Screenshot with Bounding Boxes: A screenshot of the current browser viewport with bounding boxes around clickable elements, allowing the model to decide which area to click.

This approach is based on the open-source project Browser Use.

Addressing the "Wrapper" Concern

The speaker addresses the concern that Manas is simply a "wrapper" around foundational models and may lose its edge as these models become more capable.

• Focus on Innovation Pace: The speaker argues that the key to success is the pace of innovation and the ability to quickly adapt to new models and technologies.
• Flexibility and Emergent Capabilities: Manas's simple structure allows it to leverage the best models in the world and generate emergent capabilities without requiring extensive training for specific use cases.

Local vs. Cloud Environment

Manas does not plan to offer a local environment because it believes it is important to allow users to focus on other tasks while Manas runs in the cloud. Manas plans to support virtual Windows and Android environments in the future.

Conclusion

Manas is an AI agent that leverages the power of cloud-based LLMs and a virtual computer to perform a wide range of tasks. Its core philosophy of "less structure, more intelligence" allows it to adapt to different tasks and generate emergent capabilities. By focusing on innovation and flexibility, Manas aims to remain competitive in the rapidly evolving AI landscape.