Microsoft SkillOpt: Use Agent skill documents as neural networks to train, open source in two weeks 5k+ Stars

If you have ever messed with AI Agent, you have most likely experienced this cycle: write a skill prompt → Agent runs → the effect is not good → change the prompt word → Agent runs again → better → three days later, you will find that it has become worse on another task. You gradually learn to adjust your wording intuitively, but you can't tell which sentence is better and which sentence is worse. The SkillOpt team at Microsoft Research Asia doesn't think this is right. Their question is: Why can the weights of neural networks be optimized using gradient descent, but the skill documents of Agents can only be guessed by the human brain? The answer given by SkillOpt is: the skill document itself is a trainable parameter. Don’t touch the model weights, don’t change the inference architecture, just change the skil