Beagle (DarwinX)

Salesforce AI Research's harness-evolution trainer, released as code on 2 September 2026 with the DarwinX paper from July. The premise is that an agent's capability depends on its harness (prompts, tools, skills, control flow) as much as its weights, and that single-lineage self-improvement is path-dependent: a local win often regresses other tasks. DarwinX instead treats self-evolution as selection over a population of harnesses with the model frozen. A preserve-and-extend contract admits only variants that extend coverage without regression, an archive keeps alternative lineages for recombination, and failure-, teacher-, and self-derived evidence share one edit interface; fitness comes from each benchmark's own verifier with no gold solutions. Reported results include 83.2% on Terminal-Bench 2.1 and a rise from 43.5% to 93.0% on WebArena-Infinity. The platform under the lab's co-evolving harnesses line.