JARVIS (HuggingGPT) is a collaborative system that uses an LLM as a controller to connect numerous expert AI models from Hugging Face Hub, planning tasks, selecting models, executing them, and generating responses to solve complex AI tasks from user requests.
To explore artificial general intelligence (AGI) and deliver cutting-edge research; it analyzes user requests via an LLM controller, plans tasks, selects and executes expert models hosted on Hugging Face, and integrates results to generate responses for solving complicated AI tasks.
HuggingGPT/JARVIS is an agentic system where an LLM controller autonomously plans tasks, selects expert models, and executes them in a multi-stage workflow (Task Planning → Model Selection → Task Execution → Response Generation) without per-step human confirmation. The README explicitly describes an autonomous agentic loop, and the code surface confirms it 'executes arbitrary code' and 'executes shell commands', invoking and running selected models automatically. There is no code-enforced approval gate. While the entrypoint heuristics flagged nothing, the deeper surface analysis overrides README/entrypoint and clearly shows autonomous multi-step execution with code/shell execution capabilities, placing this firmly at Tier 4.