Microsoft published Skill Factory, a module in its open-source Webwright browser-agent project that converts successful agent runs into reusable scripts. Each script takes named inputs and re-runs a web task on its own, without calling a language model, and must reproduce its recorded answers before it is added to the library. Reported gains come from tests on a small self-hosted benchmark, and the authors note that a verified script can still break when the website or operating system changes.
What changed
Agent 'skills' in comparable projects are written documents that a model must read and interpret on every run.
What it unlocks
Turning successful browser-agent runs into standalone command-line scripts that repeat the same web task without calling a model, and reusing them as a starting point when a new task is only slightly different.
- held-out accuracy 70% with library vs 55% from scratch (+15 pp)
- held-out average steps 14.7 vs 17.1
- a learned skill re-runs in about 40 seconds with no model calls
- about 40% of distillation attempts pass verification on the first try
- 7 of 30 training solves failed the correctness gate and were excluded
What you need to act on it
- clone and install the Webwright repository plus Playwright Chromium
- an OpenAI-compatible API key for every step except running an already-learned skill
- github.com2026-08-06