Salesforce Introduces DarwinX, Reporting a Browser-Agent Score Jump to 93%

Salesforce researchers report that they raised a GPT-5.5-based browser agent’s score from 43.5% to 93% without changing the model underneath it. The intervention was DarwinX: a system that evolves the prompts, tools, skills and workflows surrounding an agent, then tests whether each improvement breaks work the agent had already learned to do.

The result puts attention on the agent harness, the operating layer that tells a language model what context to use, which tools it can call and how to carry out a task. DarwinX keeps model weights frozen and modifies that surrounding layer instead, a potentially useful approach for application teams using hosted models without their own training pipelines.

A search process designed not to forget

A common self-improvement loop takes the current best agent, studies failures and patches it. Salesforce’s researchers argue that this can produce path dependence: an early, locally useful change becomes the foundation for every later change, even if another branch would have proved more valuable. It can also cause cross-task interference, where a fix for one job makes another worse.

DarwinX takes a population-based approach. It keeps multiple harness variants, retains branches with specialized strengths and can merge complementary changes. A merged candidate must still pass preservation and confirmation testing, including checks on tasks its parent versions had already solved.

The gains extend beyond one browser test

The researchers report improvements across all four benchmarks they tested, ranging from 3.4 percentage points on SWE-bench Verified to 49.5 points on WebArena-Infinity. On Terminal-Bench 2.1, a GPT-5.5-based version of Salesforce’s proprietary Monet agent rose from 75.5% to 83.2%; the evolved harness reached 84.7% with a stronger base model.

Those results do not show that extra agent effort was applied indiscriminately. On tasks both versions already completed, median turns increased from 12 to 13. On six newly solved tasks, the median rose from 11 to 22 turns, suggesting the evolved harness used more verification and retries selectively in these tests.

A promising result with an evaluation burden

DarwinX also showed a limited transfer result. A harness evolved on Terminal-Bench scored 84.2% on SWE-bench Verified, versus 80.8% for a reference harness, without receiving SWE-bench feedback during evolution. That is evidence that some learned harness behavior can travel across benchmarks, but the paper tests transfer in only one direction.

The central constraint is the evaluator. DarwinX can select a better variant only when it has a credible way to determine task success and detect regressions. That was relatively tractable in these coding and browser experiments; turning changing, ambiguous enterprise work into dependable tests remains the practical challenge behind the framework’s promise.