From 43.5 per cent to 93 per cent of browser tasks completed successfully without altering the model weights: Salesforce’s DarwinX method

Researchers at Salesforce are refining not the model itself but the framework surrounding it, by selecting from competing variants. The improvements are replicated across four separate test benches.

When an agent fails, the reflex is to accuse the model. The Salesforce AI Research team, with Ran Xu as lead author, takes the problem from the other end: it optimizes the harness—the instructions, tools and control loops that surround the model—by leaving its weights intact.

The frame, named DarwinX, proceeds by evolution: it maintains several competing variants of the harness and accepts a modification only if it improves performance without regressing on other tasks. On WebArena-Infinity, the completion rate rose from 43.5% to 93%. Terminal-Bench 2.1 grew from 75.5% to 83.2%, TerminalWorld from 61% to 68.3%, and SWE-bench Verified earned 3.4 points.

The practical interest lies in where the gain is located: in a layer that development teams already control, without access to the weights of a proprietary model. Salesforce publishes the associated tool, Beagle, on GitHub licensed under Apache 2.0.