Implementing Agent Skills: three loading decisions worth separating
ialijr · reddit · 2026-09-30
While adding skills support to a finance agent, the author found "load the skill" hides three separate architecture decisions:
- What the server loads: for a small self-maintained catalog of instruction-only skills, read and validate files once at startup into a registry—no filesystem reads per turn; edits require a restart.
- What the model sees: only names and descriptions go in the system prompt; a loadskill(name) call fetches the body on demand. Having skills in server memory doesn't mean they're in model context.
- When tools are available: the agent already owns the tools; loading a skill teaches the workflow rather than installing capabilities. Tool definitions stay stable for prompt-cache reuse, but the model can act without consulting the skill—so the author evaluates whether instructions are loaded before the actions they should inform.
The first skill skipped a code sandbox: a spending-report workflow's value lies in chart choice and query shaping, and a pure-instruction SKILL.md suffices. If your workflow executes code, host requirements change. A framework-agnostic guide with TypeScript/Python examples is linked in the comments.
More from coding & agent
- Blender MCP Optimized for Codex: View and Edit 3D Scenes Without Opening Blender — sidahuj · 2026-09-30
- This developer's perfect AI coding stack costs $420/month across ChatGPT, Claude and AmpCode — iannuttall · 2026-09-30
- A tricky UI test for agents: scrolling up to load images in long chat apps — gethackteam · 2026-09-30
- Cloudflare Birthday Week: 6x faster containers, AI Gateway auto-routing, Workers error monitoring — ritakozlov · 2026-09-30
- Tianqi Chen's team open-sources a book on compiler-driven agentic GPU kernel optimization for MLSys — sh_reya · 2026-09-30
- computesdk cuts cold-start latency 6x: median 4.05s to 648ms via co-located containers and VM restore — dinasaur_404 · 2026-09-30