Agent S4, an open-source agent scientist, enables a 27B parameter model to outperform GPT-5.5 and Opus 4.8 on computer-use tasks by autonomously evolving its own tool interfaces. Unlike fixed-interface agents that rely on screenshot-based actions, Agent S4 dynamically improves its interaction layer, reducing human engineering overhead. The system achieves this at approximately $1.10 per task while remaining fully open source.

Read original