Google Ships Agent Skills Repo to Cut AI Token Costs by 63%
Google's official Agent Skills repo ships 13 Google Cloud skills for Claude Code, Cursor, and Antigravity — solving context bloat with on-demand expertise
PRO- Google launched an official Agent Skills repository at Cloud Next 2026 with 13 skills for Google Cloud products, free and Apache 2.0 licensed.
- Skills use progressive disclosure to load only ~100 tokens of metadata at startup, pulling full instructions only when a task matches — directly solving context bloat from MCP overuse.
- Covers BigQuery, Cloud Run, Firebase, GKE, Gemini API, AlloyDB, Cloud SQL plus Well-Architected Framework pillars (Security, Reliability, Cost Optimization) and onboarding recipes.
- Evals show 96% correct API code generation with Gemini 3.1 Pro and 63% fewer tokens per correct answer when combining Skills with MCP, versus vanilla prompting.
- Works with Claude Code, Cursor, Codex, Antigravity, and Gemini CLI — install with a single
npx skills add google/skillscommand. - The format was originally open-sourced by Anthropic; Google adopting it as their official documentation delivery layer is the key industry signal here.
If you've ever wired an AI coding agent up to a documentation source and watched it gradually lose coherence after a few queries, you've hit one of the quietest and most persistent problems in agentic AI: context bloat. Google just shipped a direct answer to it. Announced at Google Cloud Next 2026, the company's official Agent Skills repository is now live on GitHub, free, Apache 2.0 licensed, and ready to drop into any agent that supports the open skills standard.
The problem nobody wanted to admit
The standard fix for agents that hallucinate outdated APIs has been to connect them to a live documentation source via MCP (Model Context Protocol , a standard for giving agents access to external tools and data). Google already runs an MCP server for its developer documentation, but heavily using MCP servers can cause "context bloat," where huge amounts of context are loaded into the context window, confusing the model and racking up token costs.
Connect a few MCP servers without managing what loads, and you quickly reach 90+ tool definitions , over 50,000 tokens of JSON schemas before the model even starts reasoning. Layer in system prompts, workflow instructions, and conversation history on top, and as context grows, the model's attention degrades and important information gets buried. The community has a name for this, and it gets worse the more products an agent is expected to know.
The real gap isn't access to information. It's the absence of condensed, agent-optimized expertise that loads on demand rather than all at once. That's the gap Agent Skills are designed to fill.
What a "skill" actually is
An Agent Skill is not a plugin, not an MCP server, and not a system prompt. It's a self-contained unit of functionality , instructions, reference documents, and assets , packaged in a folder with a SKILL.md file at its root. The clever part is how they load.
The loading mechanism follows a three-tier pattern called progressive disclosure , a term borrowed from UX design, where interfaces reveal complexity incrementally rather than dumping everything on a user at once:
- Level 1 (Discovery): At startup, agents load only the name and description of each available skill , approximately 100 tokens , no matter how rich the skill actually is.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
