Inside Google Agent Skills: How We Keep AI Agents Sharp

AI agents are only as smart as the instructions you feed them. That’s the bet behind Google Agent Skills — an open-source repo that encodes Google Cloud domain knowledge into structured, agent-readable instructions. The goal? Make coding agents smarter, safer, and more accurate, without the hallucination drama.

But here’s the thing: with 15,000+ GitHub stars and a swarm of product teams wanting in, keeping quality high while scaling is a whole different swamp. Let’s wade through how the team does it.

From Swarm to Stardom

The project didn’t start in a quiet pond. It kicked off as a fast-paced “swarm” effort leading up to Google Cloud Next 2026, with Developer Advocates and Technical Writers joining forces. The launch post dropped, and the community went wild — 15,000 stars in no time.

Soon, teams across Google (even outside Cloud, like Ads) wanted to contribute skills for their services. Popularity is great, but it brings a classic problem: how do you scale without letting quality sink?

The Quality Bar: High Enough to Bounce a Caiman

When everyone’s contributing, standards can get murky. A vague skill with broken links or missing edge cases drags down the whole agent experience. So the team set a very high bar — think basking rock under the sun, not a muddy puddle.

Every skill must pass a gauntlet of automated checks before it even thinks about merging:

• Linters: Validate frontmatter, line counts, directory layout, and naming conventions.
• Link Checkers: Test every URL to kill 404s and hallucinated links before they haunt you.
• AI-Assisted Checklists: Ensure instructions follow structural patterns and guardrails.

Evals: Not Just a One-Time Splash

Documentation changes, APIs evolve, and LLMs get upgrades. A skill that works today might flop tomorrow. That’s why the team runs continuous evaluations — both on submit and weekly.

Authors must provide evaluation prompt suites and scoring rubrics. The team measures two things: accuracy (task completion quality) and efficiency (tokens and time). They even run evals across multiple agent frameworks to get statistically solid results. The 2×2 matrix shows whether a skill actually delivers an uplift — or just paddles in place.

Skills Are Products, Not Snippets

Here’s a lesson worth swimming with: a skill is a living product, not a one-off doc. That means strict ownership rules. Repo maintainers oversee health and CI pipelines; skill owners keep their skills updated when APIs shift or quality dips.

To help authors, the team built internal skills and agentic tools (using ADK) that run multi-agent loops for authoring and self-critique. It’s like having a capybara buddy who double-checks your work before you ship.

DevRel Skills: Internal Efficiency

Beyond the public repo, there’s a parallel initiative called DevRel Skills — encoding internal workflows like content transformation, SEO, and reporting into skills. It’s a smart way to make the team’s daily grind more consistent and effective.

Dive Deeper

Want to explore the swamp yourself? Check out the Google Agent Skills repo, or read the intro guide and intermediate tutorial.

For more on building with cloud platforms, see our Google Cloud review and Supabase review for alternatives. And if you’re comparing AI model costs, our pricing comparison is a handy map.

Thanks for reading — now go build something that doesn’t hallucinate.

Original announcement published on Google Cloud.