Powered by various language models such as GPT-4, Vicuna, LLaMA, and Flan-T5

Supports multi-modality tool learning and task solving such as text, image, video and audio

Supports task decomposition into both linear task-solving plans and non-linear task-solving plans

Allows both benchmark task solving and open-ended task solving

Provides easy-to-use evaluation protocols to evaluate task-solving ability

Provide Reinforcement Learning from Task Feedback (RLTF) to allow continuously self-improving agent