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 bestskill.md reportedly transfers across models and harnesses, and improves benchmark accuracy without adding inference-time calls.

Original post →

Microsoft proposes training agent skills like neural networks, with validation gates and no weight updates

More from coding & agent