Hugging Face researchers have unveiled DarwinX, a novel approach to AI agent self-improvement that treats evolution as a selection process over a population of agent harnesses—prompts, tools, skills, and control flow—while keeping the underlying model frozen. This method addresses the limitations of single-lineage search, which often leads to path-dependent results and regressions on other tasks.
DarwinX employs a 'preserve-and-extend' contract, allowing only variants that expand coverage without causing regressions. An archive maintains alternative lineages for recombination, and failure-, teacher-, and self-derived evidence share a unified edit interface. Fitness is determined by each benchmark's own verifier, eliminating the need for gold solutions or hand-picked winners.
Across four benchmarks, DarwinX added an average of about 17 points. Notable results include Terminal-Bench 2.1 rising from 75.5% to 83.2% on GPT-5.5, surpassing Codex's 83.1%, and reaching 84.7% on a stronger base. TerminalWorld's held-out split achieved 68.3%, ahead of all off-the-shelf agents, while WebArena-Infinity real-task pass@1 jumped from 43.5% to 93.0% audit-clean. A Terminal-Bench 2.1 harness transferred unchanged to SWE-bench Verified, improving from 80.8% to 84.2%.
The researchers highlight two unexpected findings: the in-loop proxy saturates while held-out performance lags, and the best-fitting variant is not the best generalizer. This underscores the importance of maintaining a population rather than following a single best lineage. Additionally, an audit of WebArena-Infinity trajectories revealed that the gap between DarwinX and the base agent widens under scrutiny, confirming the robustness of the evolved harnesses.