Salesforce Cracks Code to Dramatic AI Agent Performance Boost

Salesforce has developed a framework called DarwinX that improves AI agent performance by 93% on a benchmark of web-based tasks. The research team, led by Yifan Zhang, Yutong Dai, and Juntao Tan, treated AI agent optimization like natural selection using the DarwinX framework.

The framework operates at the harness level, evolving the scaffolding around language models while leaving their weights untouched. This approach allows for dramatic improvements in agent reliability without expensive model retraining or risk of catastrophic forgetting.

According to the research paper published on July 31, 2026, DarwinX achieved a real-task pass rate of 93% on the WebArena-Infinity benchmark, compared to 43.5% before optimization. The framework also reduced invalid trajectories from 23.5% to 1.4%, and improved performance by an average of 17 points across all tested benchmark suites.