Microsoft proposes training agent skills like neural networks, with validation gates and no weight updates
Saboo_Shubham_ · x · 2026-07-21
Microsoft lays out a strategy for **self-evolving agent skills**: treat skills like neural-network training, with epochs, minibatch size, learning rates, and validation gates, while keeping model weights frozen. The attached SkillOpt paper/site says the system optimizes a skill document rather than model weights, using scored rollouts and bounded edits. A candidate edit is accepted only if it improves held-out validation, with learning-rate budgeting and slow meta-updates to keep training stable. The resulting `best_skill.md` reportedly transfers across models and harnesses, and improves benchmark accuracy without adding inference-time calls.
Related event: Microsoft Open-Sources SkillOpt for Self-Evolving Agent Skills(2 posts)→
More from coding & agent
- Meta and Unity link AI workflows to Quest development across setup, input and validation — Vjeux · 2026-07-21
- AI Engineer World’s Fair spotlights Kids Day with 87 children learning to code — steveonjava · 2026-07-21
- OpenCodex lets developers swap multiple model providers inside one Codex harness — arrakis_ai · 2026-07-21
- Kimi K3 looks stronger and about 5× cheaper on a frontend dashboard task — OwariDa · 2026-07-21
- This Figma MCP bridge exports real assets into your repo without API tokens or rate limits — No_Mechanic_1368 · 2026-07-21
- Hermes agents are now holding daily standups without human involvement — Teknium · 2026-07-21