DarwinX：通过自然选择进化智能体 harness
原标题：DarwinX: Evolving Agent Harnesses Through Natural Selection
AI 摘要
DarwinX 提出了一种通过自然选择进化智能体 harness（提示、工具、技能和控制流）的方法，在冻结基础模型的情况下，利用种群选择和验证器评估来提升性能。在多个基准测试中，该方法平均提升约 17 个百分点，例如 Terminal-Bench 2.1 从 75.5% 提升至 83.2%，WebArena-Infinity 从 43.5% 提升至 93.0%，且进化出的 harness 可零样本迁移至其他任务。研究表明，保持种群多样性而非单一最优路径，能有效将过拟合的代理指标转化为实际泛化收益。
正文节选
DarwinX: Evolving Agent Harnesses Through Natural Selection Abstract DarwinX evolves agent harnesses via population selection with frozen models, improving verified performance across benchmarks without benchmark-specific patches. An LLM agent's capability depends not only on model weights but on its harness: prompts, tools, skills, and control flow. Self-improvement loops already edit harnesses, yet single-lineage search is path-dependent and local wins often regress other tasks. We introduce D