InfantAgent-Next: Multimodal AI Agent for Automated Computer Interaction

A new multimodal generalist agent named InfantAgent-Next has been unveiled by researchers, aimed at facilitating automated interactions with computers through text, images, audio, and video. In contrast to current methods that depend on a singular large model or limited modularity, this agent employs a highly modular architecture that allows for the integration of tool-based and pure vision agents. This design enables various models to work together to tackle separate tasks sequentially. InfantAgent-Next showcased its versatility by performing on both vision-centric benchmarks like OSWorld, where it achieved 7.27% accuracy, outperforming Claude-Computer-Use, and on tool-heavy benchmarks such as GAIA and SWE-Bench. The evaluation scripts and code have been made publicly available.