Manus Agent Prompts
Relevant source files
The following files were used as context for generating this wiki page:
Introduction
Manus is an agentic AI system designed to operate autonomously within a Linux sandbox environment to solve complex software engineering, research, and data processing tasks. The system's behavior is governed by a set of foundational prompts that define its identity, operational constraints, and technical capabilities. Manus excels at multi-chapter report writing, application creation, and navigating the internet to gather and verify facts.
The core logic of the Manus agent relies on a structured execution loop and a modular architecture. This architecture includes specialized modules for planning, knowledge retrieval, and data sourcing, ensuring that the agent can adapt to changing requirements and troubleshoot errors independently.
Sources: Manus Agent Tools & Prompt/Modules.txt:1-15, Manus Agent Tools & Prompt/Prompt.txt:1-10
System Architecture and Modules
The Manus system is composed of several specialized modules that communicate via an event stream. This event-driven architecture allows the agent to process user messages, tool actions, and system-generated planning updates in a chronological order.
Core Modules
| Module | Description |
|---|---|
| Planner Module | Provides overall task planning using numbered pseudocode to represent execution steps. |
| Knowledge Module | Manages task-related knowledge and best practices for the agent to reference. |
| Datasource Module | Grants access to authoritative data APIs, which are prioritized over general web searches. |
| Agent Loop | The iterative cycle of analyzing events, selecting tools, and submitting results. |
Sources: Manus Agent Tools & Prompt/Modules.txt:32-75
The Agent Loop Process
The agent operates in a continuous loop to ensure tasks are completed thoroughly. This flow involves deep analysis of the event stream and strategic tool selection.
This diagram illustrates the iterative steps taken by the agent to reach a final deliverable.
Sources: Manus Agent Tools & Prompt/Modules.txt:24-33
Operational Guidelines and Rules
Manus follows strict operational constraints to maintain security, reliability, and technical accuracy. These rules cover file management, browser interaction, and coding standards.
Data and Information Priority
Manus prioritizes information sources to ensure accuracy. The hierarchy of information is:
- Authoritative data from the Datasource API.
- Original web pages accessed via a browser (snippets are considered insufficient).
- Internal model knowledge.
Sources: Manus Agent Tools & Prompt/Modules.txt:104-108
Browser and Shell Interaction
The agent utilizes a Linux sandbox (Ubuntu 22.04) with full internet access. When interacting with the web, the agent must access original URLs rather than relying on search result snippets. For shell operations, Manus is instructed to use non-interactive commands (e.g., using -y or -f flags) to avoid blocking the execution loop.
The sequence shows the agent's requirement to verify information by accessing original sources.
Sources: Manus Agent Tools & Prompt/Modules.txt:110-135
Writing and Coding Standards
Manus is optimized for high-volume, detailed output. Technical requirements include:
- Continuous Prose: Use paragraphs instead of lists unless requested.
- Detailed Citations: Every significant piece of information must be cited with original sources and URLs.
- Code Execution: Code must be saved to files before execution; direct input to interpreters is forbidden.
- Validation: For mathematical calculations, Manus must use Python or
bcrather than mental calculation.
Sources: Manus Agent Tools & Prompt/Modules.txt:137-160
Toolset and Capabilities
Manus is equipped with a comprehensive suite of tools for navigating both the local filesystem and the wider web.
Tool Categories
| Category | Capabilities |
|---|---|
| Filesystem | Reading, writing, appending, searching, and compressing files. |
| Web | Searching, navigating, screenshotting, and interacting with UI elements. |
| Development | Shell access, package installation, and port exposure for deployment. |
| Data API | Specialized Python-based ApiClient for accessing curated data banks. |
Sources: Manus Agent Tools & Prompt/Prompt.txt:15-55, Manus Agent Tools & Prompt/Modules.txt:68-75
Example: Data API Usage
Data APIs are accessed through Python scripts rather than direct tool calls, allowing for complex data manipulation.
import sys
sys.path.append('/opt/.manus/.sandbox-runtime')
from data_api import ApiClient
client = ApiClient()
# Retrieving weather data via the WeatherBank API
weather = client.call_api('WeatherBank/get_weather', query={'location': 'Singapore'})
print(weather)
Sources: Manus Agent Tools & Prompt/Modules.txt:77-85
Task Management and Progress Tracking
The agent manages its workflow using a todo.md file, which serves as a checklist derived from the Planner module's high-level pseudocode.
- Checklist Creation: A
todo.mdis generated at the start of a task. - Immediate Updates: Markers in the checklist are updated via text replacement as soon as a sub-task is finished.
- Rebuilding: If the overarching task objective changes significantly, the agent rebuilds the
todo.md. - Verification: Before task completion, Manus verifies that all planned steps are reached and redundant items are removed.
Sources: Manus Agent Tools & Prompt/Modules.txt:87-95, Manus Agent Tools & Prompt/Agent loop.txt
Conclusion
Manus Agent Prompts establish a highly structured, autonomous environment for AI-driven task execution. By enforcing an iterative agent loop, clear information hierarchies, and modular planning, the system ensures that complex requests are handled with technical precision and transparency. The integration of specialized modules like the Planner and Datasource API enables the agent to act as a comprehensive technical partner capable of end-to-end task completion.