Mentis is a Python framework for building and orchestrating multi-agent systems, leveraging LangGraph for state-driven planning and execution. It targets developers and researchers building complex AI applications, enabling sophisticated task automation through collaborative agents.

How It Works

Mentis employs a state-based planning architecture where a Supervisor Agent coordinates Specialist Agents. A dedicated Planner node creates an initial plan, which the Supervisor then executes by delegating tasks to specialized agents (e.g., Coder, Researcher). An Evaluator node assesses agent outputs, updating the plan's state. This iterative process, managed by LangGraph's graph execution and checkpointer for persistence, allows for complex, multi-step task resolution and inter-agent communication via the A2A protocol.

Quick Start & Requirements

• Installation: uv sync or pip install -r requirements.txt after setting up a Python 3.11+ environment.
• Prerequisites: API keys for LLMs (OpenAI, DeepSeek, XAI Grok), search tools (Tavily, Exa), and optional tools like E2B or Replicate. LangSmith keys are highly recommended for debugging.
• Running Examples: Execute example scripts like python examples/state_based_supervisor_examples/03_multi_agents.py from the project root.
• Documentation: Project structure and configuration details are in the README.

Highlighted Details

• State-driven planning with dedicated Planner, Supervisor, and Evaluator nodes.
• Support for Google's Agent-to-Agent (A2A) protocol for interoperability.
• Modular agent design with BaseAgent and ReactAgent, facilitating custom agents.
• Centralized tool registration and dynamic loading.
• Persistence via LangGraph's Checkpointer mechanism.
• Includes a "Super Agent" concept for complex, end-to-end tasks, exemplified by the DeepResearch Agent.

Maintenance & Community

The project is actively developed, with contributions welcomed via Issues and Pull Requests. Contact information (WeChat) is provided for support.

Licensing & Compatibility

Licensed under the MIT License, permitting commercial use and integration with closed-source projects.

Limitations & Caveats

The requirements.txt file is noted as not actively maintained, requiring users to ensure all necessary libraries are included. Future work includes enhancing agent toolsets, evaluator logic, and task dependency handling.