Google's adk-recipes repo sorts agent-building examples into three layers

One repo replaces the old samples

As developers.googleblog.com reported, Google has restructured its Agent Development Kit (ADK, its toolkit for building AI agents) examples into a repository called google/adk-recipes. It succeeds google/adk-samples and now feeds Google's Agent Platform from a single curated source.

The notable design choice is the intended reader. Google says every piece was built to be machine-readable first, so a coding assistant can discover a relevant recipe for a use case, then install and use it with little hand-holding. Human-facing README files still exist, but they are no longer the main interface.

Three folders, three jobs

The top level is split into core/, plugins/ and contrib/. Core holds reference patterns curated by Google's Agents CLI and developer-relations teams. Plugins holds installable industry skills. Contrib holds complete services you can deploy, written in five languages.

Every recipe carries a validated manifest.yaml, which is a structured description of what the recipe is. It also has an AGENTS.md, a plain note telling the assistant what the recipe does and how to use it. Contrib and plugins recipes add a Dockerfile (a packaging recipe for running the code) and a test that checks it actually runs.

Automated checks enforce these standards on every pull request. The repo also includes assistant skills that handle most of the preparation, so an agent can scaffold, validate and submit a recipe itself.

core/: a harness for agents that last days

The featured core example is Horizon, found at core/python/long-horizon-harness. It targets a hard problem: an agent that stays useful across days, remembering what a user did yesterday and planning for tomorrow. The source lists the recurring pieces as memory across sessions, isolating each user's tool use in a sandbox, managing secrets, guardrails against runaway behavior, and steps that can resume after interruption.

Horizon bundles those pieces in one place. It can run locally with in-memory sessions and no cloud deployment, and a separate command rolls it out to Cloud Run through Terraform. Its test suite has eighteen named rubrics covering memory recall, attempts to leak data through injected instructions, guardrails under adversarial input, and more.

Google's advice is that when you ask an assistant to build something that must run across days, it should start from Horizon as the scaffold.

plugins/: retail skills you install in one line

Plugins are packaged under the Agent Plugins v1.0.0 specification, an open standard for installable AI skills. A skill is a folder with an instruction file (SKILL.md), a README, quality checks (EVAL.yaml) and supporting scripts. Retail is the first industry fully covered, with more promised as their own folders.

The retail-product-search skill builds an agent that finds products by meaning rather than exact keywords. It loads a catalog into BigQuery (Google's data warehouse), creates embeddings (numeric fingerprints of meaning), builds a Vector Search collection and deploys to Cloud Run. If its quality checks fall under an 80 percent passing target, it stops and reports the failures. In the demo, a five-product catalog loaded with no errors and a query for wireless headphones under 200 returned results.

The retail-virtual-tryon skill creates a realistic image of a model wearing or holding a product. It can add an eight-second catwalk video made with Veo 3.1. By default it uses gemini-2.5-flash-image, switching to gemini-2.5-pro-image when quality matters more than speed.

Both install with an npx skills command into many assistants, including Claude Code, Cursor and Antigravity. The tool's own output warns that installed skills run with full agent permissions, so review them first.

contrib/ and what is still unclear

Contrib is where recipes become deployable services. Twelve Python agents live there today, spanning finance, evaluation, document and knowledge pipelines, brand and content work, and retail. The source text we have cuts off partway through this section, so we cannot describe the full list or how mature each agent is.

The path into the public gallery is also worth knowing. A pull request to contrib puts a recipe on GitHub where others and their assistants can find it. Appearing in Agent Garden, Google's gallery, is a separate editorial decision by that team, which then allows one-click installs inside a Google Cloud project.