DarwinX evolves agent harnesses via natural selection, adding ~17 points per loop
bendee983 · x · 2026-09-18
A new arXiv paper, DarwinX: Evolving Agent Harnesses Through Natural Selection, treats agent self-improvement as selection over a population of harnesses with the model frozen, fixing path dependence and cross-task interference of single-lineage search. A preserve-and-extend contract admits only non-regressing variants; an archive keeps lineages for recombination; fitness comes from each benchmark's own verifier. Results: +17 points per evolution loop on average; Terminal-Bench 2.1 rises +7.7 to 83.2% (84.7% on a stronger base); WebArena-Infinity pass@1 jumps 43.5%→93.0%; a harness transfers unchanged to SWE-bench Verified. Code for Beagle is open-sourced.
Related event: Salesforce Unveils DarwinX: Evolving Agent Harnesses via Natural Selection(2 posts)→
More from coding & agent
- Interactive 404 page built with one prompt and Opus: welcome to design engineering — Tegadesigns · 2026-09-18
- Offline replay of 3,300 production decisions validated Jev routing in an afternoon for 20 cents — mikegiannulis · 2026-09-18
- Same Material, Better Order: Reranking Lifts Correct-Passage Hit Rate to 90% — mikegiannulis · 2026-09-18
- 13 'Please Text Me, No Calls' Messages Misrouted — Small Model Caught All of Them — mikegiannulis · 2026-09-18
- Intent Routing 90.5% vs 83.8%, 'Already Did That' Detection 98.3% vs 96.6% — mikegiannulis · 2026-09-18
- 95% Accuracy, 0.057 Brier Score on Faithful-vs-Fabricated Pairs — With Caveats — mikegiannulis · 2026-09-18