DarwinX proposes evolving the harness around a frozen LLM — its prompts, tools, memory, and control flow — using a population-based natural-selection process rather than retraining the underlying model. The system maintains an archive of harness variants and only promotes a change when it extends capability on some tasks without causing regressions on others. Tested on four benchmarks with increasing distribution shift between training and evaluation, DarwinX lifted Terminal-Bench 2.1 accuracy from 75.5% to 83.2% on a frozen GPT-5.5 model, raised WebArena-Infinity audit-clean pass@1 from 43.5% to 93.0%, and showed a harness evolved on Terminal-Bench transferring unchanged to reach 84.2% on SWE-bench Verified.
