Google Paper Formalizes Agent Skills Infrastructure

The paper’s real contribution is treating skills as infrastructure—closer to governed packages and workflow operators than prompt snippets. That framing is useful, but it also exposes a hard problem: reusable agent capabilities amplify retrieval errors, procedural drift, and supply-chain attacks.

• –A six-tuple model distinguishes skills from prompts, tools, workflows, and memory through activation conditions, constraints, interfaces, policies, and state effects.
• –Capability-based permission gates and sandboxed state transitions offer a practical foundation for limiting what a skill can execute.
• –Dynamic routing becomes a core systems challenge as registries grow; flat semantic search can surface irrelevant or conflicting procedures.
• –Natural-language instructions and activation conditions remain difficult to verify statically, leaving room for prompt injection and malicious skill packages.
• –This is a reference architecture and research synthesis, not a shipped runtime; developers will still need interoperable formats, trustworthy registries, and realistic evaluations.