DarwinX

DarwinX is a framework, described in a preprint, for evolutionary self-improvement of LLM agents. Instead of tuning model weights, it evolves the agent's harness — prompts, tools, skills and control flow — treating self-evolution as selection over a population of harnesses with the model frozen. Key mechanisms include a "preserve-and-extend contract" (admitting only variants that extend coverage without regression), an archive that maintains alternative lineages, and a unified evidence interface. Fitness comes from benchmark verifiers rather than gold solutions.

DarwinX addresses task regression in single-lineage self-improvement by enabling multi-path evolution that adds capability without losing existing coverage. Results from the abstract: Terminal-Bench 2.1 — 83.2%, TerminalWorld — 68.3%, WebArena-Infinity — 93.0% (pass@1), with an average improvement of ~17 points per evolution loop; evolved harnesses transfer across benchmarks and base models. The preprint was released on July 31, 2026 (Salesforce AI Research), and the official implementation is open-sourced within the Beagle platform.