Salesforce's DarwinX framework evolves agent harness, lifting browser-task completion to 93%

Salesforce AI Research and the Salesforce Agentforce team have introduced DarwinX, a new framework that applies an evolutionary approach to optimizing an AI agent's prompts, tools, skills, and workflows — collectively called the harness — without modifying the underlying model weights. On WebArena-Infinity, DarwinX raised the agent's completion rate on 1,260 unseen browser tasks from 43.5% to 93%, with gains across all four benchmarks tested.

The researchers said that repeatedly rewriting a single harness tends to create path dependence and cross-task interference, where an edit that helps one task can quietly break others. DarwinX frames the problem as one of selection: it generates multiple harness variants, stores them in an archive, and only allows changes to advance when they improve capability while keeping regressions on previously solved tasks within a bounded tolerance. The framework also separates exploration from confirmation, rerunning promising edits at higher fidelity before they shape future generations.

In experiments with Salesforce's proprietary Monet agent, DarwinX increased Terminal-Bench 2.1 scores from 75.5% to 83.2% on frozen GPT-5.5, and raised SWE-bench Verified results from 80.8% to 84.2% without receiving SWE-bench feedback during evolution. The evolved harness added skills such as defining expected outcomes, checking generated files and grounding outputs in real tool execution. The researchers also showed that specialist variants from different branches can be merged into a stronger harness.

Because the approach leaves model weights untouched, the team says it is relevant for developers building on hosted models. Salesforce has also released Beagle, an open-source infrastructure framework for experimenting with agent evolution, under the Apache 2.0 license on GitHub.