Salesforce AI Research has introduced DarwinX, a framework that optimizes AI agents by evolving their scaffolding—including prompts, tools, and control flow—without modifying the underlying model weights. The approach uses population-based natural selection to refine agent configurations across 1,260 tasks.
In testing on the WebArena-Infinity benchmark, the system increased the task pass rate from 43.5% to 93.0%. Additionally, the rate of invalid agent outputs decreased from approximately 23.5% to 1.4%. The research team, led by Yifan Zhang, Yutong Dai, and Juntao Tan, published their findings on July 31, 2026.
On the Terminal-Bench 2.1 benchmark, DarwinX improved performance by 7.7 points to reach 83.2%. When applied to a more powerful model, the score rose to 84.7%. Across all tested suites, the framework demonstrated an average improvement of around 17 points, with successful transferability between different benchmarks.
The 12-author team utilizes a preserve-and-extend contract to ensure that only performance-enhancing changes are retained. This method avoids the high compute costs and risks associated with fine-tuning large language models, potentially offering Salesforce a competitive advantage for its Agentforce platform.
DarwinX is currently a research initiative. The published paper, identified as 2608.07545 on arXiv, does not provide details regarding production deployment, and the results have not yet undergone independent verification.
