TaskGen is an open-source, task-based agentic framework designed for efficient LLM agent execution. It targets developers and researchers building complex agentic workflows, offering a structured alternative to conversational LLM frameworks by leveraging StrictJSON for precise output parsing and Chain-of-Thought prompting.

How It Works

TaskGen utilizes StrictJSON, a JSON parser with type checking, as its core communication protocol between LLM agents. This approach enables agents to naturally perform Chain-of-Thought reasoning by using JSON keys and descriptions as guides, leading to more efficient and less verbose interactions compared to free-text conversational frameworks. Agents can be structured into single agents with LLM or external functions, or meta-agents with inner agents, supporting features like shared variables for multi-modality, RAG over function spaces, and memory.

Quick Start & Requirements

• Install via pip: pip install taskgen-ai==3.3.4
• Requires an LLM provider (e.g., OpenAI) and API keys.
• Recommended LLMs for robust performance: gpt-4o-mini, gpt-4o, Llama 3 70B. Weaker models like gpt-3.5-turbo may require more explicit prompting.
• Official Tutorials: [Link to Tutorials, if available in README]

Highlighted Details

• StrictJSON core for structured, type-checked agent communication.
• Supports single agents, meta-agents with inner agents, and RAG over function spaces.
• Features shared variables for multi-modal data and global context for persistent agent state.
• Includes an AsyncAgent for parallel task execution.
• Community uploading and downloading of agents and functions.

Maintenance & Community

• Original creator John Tan Chong Min has forked the project to his own repository.
• Key contributors include Prince Saroj, Hardik Maheshwari, Bharat Runwal, Brian Lim, and Richard Cottrill.
• Mentors/Funders include Ambuj Kumar, Alankrit Chona, Mehul Motani.

Licensing & Compatibility

• The README does not explicitly state a license. Compatibility for commercial use or closed-source linking is undetermined.

Limitations & Caveats

• gpt-3.5-turbo exhibits limitations with mathematical functions and memory features; gpt-4o-mini or better is recommended.
• The framework is under active development, and extensive testing across various LLMs and datasets is encouraged. Prompt engineering may be required for non-OpenAI LLMs.