I've been thinking about how we structure "skills" in agent systems. Across different frameworks, "skills" can mean very different things: a tool / function a role or persona a multi-step workflow But there are actually two separate questions here: What does the skill describe? persona tool workflow How does it execute? stateless (safe to retry, parallelize) stateful (has side effects, ordering matters) Most frameworks mix these together. That works fine in demos — but starts to break in real systems. For example: a tool that reads data behaves very differently from one that writes data a workflow that analyzes is fundamentally simpler than one that publishes results Once stateful steps are involved, you need more structure: checkpoints explicit handling of side effects sometimes ...
No Comments 
