My team has recently upgraded Copilot SDK from 0.3.x to 1.0.x for running a few fixed scenarios to track the agent behavior with our skills. We give the agent a generic documentation tool that searches in MS Learn and skills with service oriented instructions. We noticed that after upgrading to Copilot SDK 1.0.x, the agent shows a much higher likelihood to use documentation tools instead of loading a skill. This is not an ideal behavior. Local skills are usually knowledge the user creates to override the generic knowledge in the training data or public documentation. They are also usually more efficient in token usage.
My team has recently upgraded Copilot SDK from 0.3.x to 1.0.x for running a few fixed scenarios to track the agent behavior with our skills. We give the agent a generic documentation tool that searches in MS Learn and skills with service oriented instructions. We noticed that after upgrading to Copilot SDK 1.0.x, the agent shows a much higher likelihood to use documentation tools instead of loading a skill. This is not an ideal behavior. Local skills are usually knowledge the user creates to override the generic knowledge in the training data or public documentation. They are also usually more efficient in token usage.