DarwinX: Evolving Agent Harnesses Through Natural Selection

A recent study published on arXiv (ID: 2608.07545) presents DarwinX, a novel technique for evolving LLM agent harnesses—encompassing prompts, tools, skills, and control flow—using principles derived from natural selection. This method focuses on self-evolution by selecting from a population of harnesses while keeping the model static. Key features include a preserve-and-extend contract that allows only variants that enhance coverage, an archive for alternative lineages to facilitate recombination, and a unified editing interface for various types of evidence. Fitness evaluation is conducted by each benchmark's verifier, eliminating the need for gold standards. Across four benchmarks, DarwinX achieves an average increase of approximately 17 points, with a notable +7.7 improvement on Terminal-Bench 2.1, reaching 83.2% on a matched base and 84.7% on a stronger one. The study also mentions findings on TerminalWorld, though specific figures are absent in the abstract. This research addresses the challenges of single-lineage searches in self-improvement loops, which can lead to local optima that hinder other tasks. By utilizing a population and archive, DarwinX seeks to mitigate these limitations, contributing to advancements in AI and machine learning, especially in agentic systems and self-enhancement.