agent-skill-evolutionlisted

**CONSUMES**: Agent description + self-improvement requirements + optional existing skill/memory docs + evaluation setup (held-out set, success metric) **PRODUCES**: Applied self-evolution judgment rules + architecture decision + training loop design + safety mechanism review + gate configuration guidance # Agent Skill Evolution Capability Pack **Version**: 0.1.0 **Compatibility**: Claude Code / Codex / Copilot / OpenClaw (via SkillOpt-Sleep adapters) **License**: Apache 2.0 --- ## What This Pack Does AI agents asked to "learn from experience" or "improve over time" default to vague suggestions: "add a feedback loop," "fine-tune the model," or "log failures and iterate." These miss the engineering reality: an agent modifying its own instructions without a validation gate, edit budget, and protected regions is not learning — it is drifting. SkillOpt's empirical evidence: ungated self-modification collapsed accuracy from **0.554 to 0.026 (−52.8 percentage points)** over 5 nights. This pack embeds the judgment rules from SkillOpt (Microsoft, arXiv 2605.23904) — the first production-grade framework for text-space skill optimization, proven across 6 benchmarks, 7 models, and 3 harnesses (all 52 cells best/tied-best, with improvements of **+23.5, +24.8, +19.1 pts** on key benchmarks). The rules cover architecture decisions, training loop design, edit safety, validation gates, offline consolidation (sleep cycles), and multi-timescale memory. **Pack = self-evolution judgment.