Stanford's DarwinX evolves AI agents without retraining
New paper shows natural selection improves AI agents 17 points across 4 benchmarks...
Researchers from Stanford University and collaborators have published a paper introducing DarwinX, a novel framework that treats AI agent self-improvement as natural selection over a population of 'harnesses'—combinations of prompts, tools, skills, and control flows. Unlike traditional approaches that edit individual harnesses, DarwinX maintains an archive of alternative lineages and recombines successful variants while preserving core capabilities.
Key to DarwinX is its fitness evaluation system, which uses each benchmark's own verifier to score variations without gold solutions or hand-picked winners. The method achieved remarkable gains: Terminal-Bench 2.1 improved by 7.7 points to 83.2%, WebArena-Infinity's real-task pass@1 jumped from 43.5% to 93.0%, and the approach generalized across different tasks and models. Crucially, the evolved harnesses transferred unchanged to SWE-bench Verified, demonstrating general agent competence rather than benchmark-specific optimizations.
- DarwinX evolves AI agents through natural selection of harness variants without retraining the base model
- Achieved +17 points average improvement across Terminal-Bench, WebArena, and SWE-bench with frozen models
- Uses benchmark-specific verifiers for fitness evaluation, enabling recombination of successful lineages
Why It Matters
Turns evaluation compute into durable capability gains, enabling AI agents to improve continuously without expensive retraining cycles