TanStack AI adds lazy-loaded agent skills
TanStack’s withSkills() middleware replaces bloated system prompts with a compact skill catalog and a load_skill tool, loading full instructions only when needed. It works with any tool-calling model and supports inline, filesystem, static, and custom skill sources.
This is a practical fix for context-window waste, but lazy loading trades prompt size for an extra tool call and another opportunity for model routing errors.
- –Keeps every request’s baseline context small as skill libraries grow
- –Provider-agnostic middleware avoids locking skill delivery to one model vendor
- –Custom sources make database-, S3-, and registry-backed skill libraries first-class
- –Resource loading extends the pattern beyond instructions to references and assets
- –Developers still need clear skill descriptions and sensible access controls
DISCOVERED
9d ago
2026-08-28
PUBLISHED
9d ago
2026-08-28
RELEVANCE
AUTHOR
tan_stack