Implementing Agent Skills support: three loading decisions worth separating
retained text
While adding skills support to my finance agent, I found that “load the skill” hides three separate architecture decisions: **1. What does the server load?** For a small catalog of instruction-only skills I maintain myself, I read and validate the files once at startup and keep them in a registry. That avoids filesystem reads on every turn. The files remain the authoring format, and edits require a restart. **2. What does the model see?** Only names and descriptions go into the system prompt. A `load_skill(name)` call retrieves the body when the task needs it. Having every skill in server memory doesn't mean every skill is in model context. **3. When are the tools available?** My agent already has the tools the skill refers to. Loading the skill teaches the workflow; it doesn't install capabilities. That's manageable for a small tool set and keeps definitions stable for potential prompt-cache reuse. A large catalog may justify tool discovery too. This last choice has a consequence: the model can attempt the work without consulting the skill. So a tested loader isn't enough. I evaluate whether the instructions are loaded before the action they should inform, and whether the agent avoids unnecessary charts for simple questions. I also chose not to add a code sandbox for the first skill. It's a spending-report workflow, and the value is in choosing a useful chart and shaping the query. Tested UI components render the result. SKILLmd alone is enough for an instruction-only skill; scripts and other resources are optional. If your workflow depends on executing code, that changes the host requirements. I wrote a framework-agnostic guide around these decisions, with TypeScript/Python examples and Cameron. The link will be in the comments. If you're building a skills system, where does most of the value come from in your agent: reusable instructions, executable scripts, or access to supporting resources?
document fetched 2026-10-01 00:20 · canonical https://reddit.com/r/AI_Agents/comments/1wu4jtz/implementing_agent_skills_support_three_loading