Agent Skills are loaded in one prompt, with caching reducing multi-turn token cost
dotey · x · 2026-07-26
Skills in an agent workflow are loaded once as part of the system prompt, not through a separate model call.
- The model sees all Skill names and descriptions at the start of the conversation.
- In a single call, it matches the user request to the right Skill, reads the corresponding SKILL.md, and decides whether to call tools.
- The expected token overhead is mitigated by prompt caching: the system prompt, including Skill metadata, is reused across turns so later messages do not pay the same cost again.
More from coding & agent
- A better agent feedback loop captures clicks, traces, and console logs — jasonkneen · 2026-07-26
- Cindy launches as an open-source AI agent client for real software tasks — yangyi · 2026-07-26
- Engineers are using Grok Build voice input to drive workflow commands — yunta_tsai · 2026-07-26
- A developer asks whether AI coding tools are helping them learn or think less — Terrible_Spare_8371 · 2026-07-26
- Plan review can speed up code review, but it cannot replace it — zeeg · 2026-07-26
- Builder ships SHADDAI, a 7-agent AI team for research, writing, images and security checks — Enough-Desk3317 · 2026-07-26