SkillOpt: Agent skills as trainable parameters
| Source: Microsoft Research Blog
Tags: SkillOpt, Microsoft Research, AI agents, agent skills, prompt optimization, LLM reliability
Microsoft Research's SkillOpt treats agent instruction files as trainable parameters outside a frozen model, running a text-space forward-backward-update loop that tops all 52 evaluation cells across six benchmarks, seven models, and three execution modes — without updating any model weights.
Details
Reliable AI agents depend on well-written instruction files ('skills'), but today those files are crafted by hand or regenerated in one shot, with no principled optimization. SkillOpt reframes skill writing as a training problem: the target model is frozen, and a separate optimizer model iteratively edits the skill file through a forward (task execution) and backward (trajectory reflection) cycle, with validation gating to reject harmful revisions.\n\nThe system keeps skills compact and auditable through bounded text edits, rejected-edit feedback, and slow/meta updates that prevent prompt drift. SkillOpt is evaluated across six benchmarks, seven target models, and three execution modes — 52 evaluation cells total — and achieves the best or tied-best result in every single one.\n\nTransfer is a key result: the optimized skills generalize across model scales, agent harnesses, and related tasks, suggesting they capture reusable workflow knowledge rather than dataset-specific tricks. For practitioners building production agents, this offers a systematic alternative to hand-tuning prompts one revision at a time.