Google restructures ADK samples into adk-recipes, built for coding agents

Three folders, three reader actions

The repository is the restructured successor to google/adk-samples. Google says it splits content into three top-level folders. The core/ folder holds reference patterns curated by the Agents CLI and DevRel teams. The plugins/ folder contains installable vertical skills packaged as Plugin Recipes that conform to the Agent Plugins v1.0.0 spec. The contrib/ folder holds end-to-end deployable services across five languages.

Every recipe ships a validated manifest.yaml and an AGENTS.md that tells a coding assistant what the recipe is and how to use it. Recipes in contrib/ and plugins/ also include a Dockerfile and a runnability test. According to the authors, CI enforces these standards on every pull request, and a set of coding-assistant skills in the repo, such as prepare-python-recipe, generate-manifest and github-pr-review, does most of the preparation so an agent can scaffold, validate and submit a recipe end to end.

The post states that every AGENTS.md, manifest.yaml and SKILL.md is written primarily for the AI coding assistant that opens the repo on a developer's behalf.

What it powers on Agent Platform

Google describes adk-recipes as the source for three Agent Platform surfaces, each reading the repository in its own way.

• Agents CLI reads core/. Running uvx --from google-agents-cli agents-cli create my-agent can scaffold a core/ recipe on demand. The post says the CLI launched earlier this year at Google Cloud Next and covers a scaffold, deploy, evaluate and publish lifecycle.
• Agent Garden features a curated selection. Developers contribute through a pull request to contrib/, after which a recipe becomes discoverable on GitHub. Appearing in the Agent Garden gallery is a separate editorial decision by the Agent Garden team. Once an agent is listed, any developer in a Google Cloud project gets a one-click install.
• Gemini Enterprise app is also named as a consumer of the repository.

Google Cloud components also appear inside the recipes. Horizon uses Agent Platform Sandboxes for per-user sandboxing, Memory Bank for durable memory and Agent Platform Sessions for session state. For assistants outside Google tooling, including Claude Code, Antigravity, Cursor, Windsurf and Zed, Plugin Recipes can be installed through the open npx skills CLI maintained by vercel-labs.

Horizon: a long-horizon reference harness

The core/ example is Horizon, located at core/python/long-horizon-harness. Google says it collects the primitives that agents running across days of interaction need: cross-session memory, per-user sandboxing of tool execution, secrets management, guardrails against runaway behaviour, and orchestration that makes multi-step tasks resumable.

The default setup uses a local environment backend with in-memory sessions, so no cloud deployment is needed to start with make dev-local. The command make deploy rolls the backend to Cloud Run through Terraform, and make destroy tears it down. Horizon's evaluation suite sits in tests/eval/evalsets/ as eighteen named behavioural rubrics. They cover memory recall, injection-based exfiltration attempts, guardrail halting under adversarial input, tool-selection discipline and a journey-level plot-artifact evaluation, and run with uv run pytest tests/eval.

Retail Plugin Recipe: the vertical layer

The second layer holds industry-specific Plugin Recipes. Retail, under plugins/retail/, is the first vertical fully covered at launch. The format builds on the Agent Plugins v1.0.0 specification published at agent-plugins.org, which defines a plugin as a bundle of skills with a top-level plugin.json manifest declaring name, version, license and ownership. Each skill has a SKILL.md conversational installer, a README.md, an EVAL.yaml of quality gates, and supporting scripts and assets.

Two skills ship today. retail-product-search builds an ADK agent with a vector-search retrieval tool backed by Vector Search 2.0 on Gemini Enterprise Agent Platform and a BigQuery catalog. It handles ingesting a CSV or JSON catalog, generating embeddings with Vertex AI Embeddings, building the Vector Search collection, scaffolding the agent with a retrieve_docs tool and deploying to Cloud Run. The agent is deliberately a single root_agent with one retrieval tool.

retail-virtual-tryon generates a photorealistic composite of a model wearing or holding a product, with an optional eight-second studio catwalk video. It uses gemini-2.5-flash-image by default and falls back to gemini-2.5-pro-image when editorial quality matters more than latency.

What it means for developers

For developers on Google Cloud, the change puts reference patterns, vertical skills and deployable services in one curated tree that both people and coding assistants can search and install from. Because the same repository feeds Agents CLI, Agent Garden and Gemini Enterprise app, a contribution made through a pull request to contrib/ becomes part of the public ADK ecosystem. Google positions the machine-readable design as the way for an agent to discover, install and use recipes from a developer's preferred harness.