Agent task execution on SkillsBench 1.0 (test)
71.1Pass Rate (With Skills)EvoSkills (Claude Opus 4.6)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| EvoSkills (Claude Opus 4.6)Skill evolution source=Self-evolved, Agent harness=Claude-Code, Evolution stage parameters=K=5 oracle interventions, M=15 surrogate retries, Evaluation stage timeout=7200s/task, Backbone=Claude Opus 4.62026.04 | 71.1 | 30.6 | 40.5 | |
| EvoSkills (GPT-5.2)Skill evolution source=Self-evolved, Agent harness=Codex, Evolution stage parameters=K=5 oracle interventions, M=15 surrogate retries, Evaluation stage timeout=7200s/task, Backbone=GPT-5.22026.04 | 69.8 | 29.6 | 40.2 | |
| EvoSkills (GPT-5.2)Skill evolution source=Cross-Model Transfer (Opus 4.6 Evolved Skills), Agent harness=Codex, Evaluation stage timeout=7200s/task, Backbone=GPT-5.22026.04 | 65 | 29.6 | 35.4 | |
| EvoSkills (Claude Sonnet 4.5)Skill evolution source=Cross-Model Transfer (Opus 4.6 Evolved Skills), Agent harness=Claude-Code, Evaluation stage timeout=7200s/task, Backbone=Claude Sonnet 4.52026.04 | 63.1 | 20 | 43.1 | |
| EvoSkills (Claude Haiku 4.5)Skill evolution source=Cross-Model Transfer (Opus 4.6 Evolved Skills), Agent harness=Terminus-2, Evaluation stage timeout=7200s/task, Backbone=Claude Haiku 4.52026.04 | 54.5 | 10.4 | 44.1 | |
| EvoSkills (Qwen3 Coder)Skill evolution source=Cross-Model Transfer (Opus 4.6 Evolved Skills), Agent harness=Terminus-2, Evaluation stage timeout=7200s/task, Backbone=Qwen3 Coder2026.04 | 50.8 | 8.4 | 42.4 | |
| EvoSkills (DeepSeek V3)Skill evolution source=Cross-Model Transfer (Opus 4.6 Evolved Skills), Agent harness=Terminus-2, Evaluation stage timeout=7200s/task, Backbone=DeepSeek V32026.04 | 48.8 | 13 | 35.8 | |
| EvoSkills (Mistral Large 3)Skill evolution source=Cross-Model Transfer (Opus 4.6 Evolved Skills), Agent harness=Terminus-2, Evaluation stage timeout=7200s/task, Backbone=Mistral Large 32026.04 | 43.1 | 4.9 | 38.2 |