From Static to Dynamic Skills: A Different Model for Agent Knowledge

Current AI agent skills are often static, freezing facts and procedures together in files, which quickly become stale and contradictory. This approach is flawed because it acts like a cache with no invalidation protocol, leading to "skill inflation" as developers copy and modify outdated information. The proposed solution is to separate a skill's intent and procedure from the factual knowledge it uses by defining a scope over a live knowledge base. In this dynamic model, facts are resolved in real-time at the moment of the agent's request, generating a temporary context document for each call. This composition process uses a graph of strategies to select relevant information and adheres to a token budget by ranking content sections, ensuring the agent always has the most current context.