Yadong Lu
@Adamlu28
Computer use agent, MTS @ Microsoft
Joined February 2015
65 Following 42 Followers
I’ve always been bothered by how much time agents spend reading skill files and figuring out the exact commands to run.
Skill files are essentially lossy compression of day-to-day workflows. When invoked in Codex or Claude, the agent spend reasoning tokens “decompressing” them back into actions, adding latency and cost.
What if we traded a bit more upfront compression for nearly token-free decompression—turning user workflow trajectories directly into executable programs?
Check out our early exploration by @demisama_.
Show more
Excited to share my first project @MSFTResearch 🎉!!
We built Web Skill Factory, which turns successful web-agent runs into reusable programs: verify what worked, distill it into parameterized code, then run or adapt it for future tasks.
On a held-out WebArena subset, skill reuse improved accuracy from 55% to 70% and reduced average steps from 17.1 to 14.7. A fresh agent solve can take around 10 minutes, while replaying a skill directly takes about 40 seconds with zero tokens.
Our goal is to build skills agents can execute, not just instructions they reread. Feedback is very welcome, and a GitHub star would mean a lot!
🔗 Check it out:
A few more thoughts below 🧵
Show more
