Google's adk-recipes repo splits ADK examples into core, plugins and contrib layers

The repository has three top-level folders. core/ holds reference patterns curated by the Agents CLI and DevRel teams, plugins/ holds installable vertical skills packaged as Plugin Recipes that follow the Agent Plugins v1.0.0 spec, and contrib/ holds end-to-end deployable services in five languages. Each recipe includes a validated manifest.yaml and an AGENTS.md written for coding assistants, and CI checks the standards on every pull request.

Horizon, a long-horizon harness at core/python/long-horizon-harness, is the reference pattern in core/. It covers cross-session memory, per-user sandboxing and guardrails, and its evaluation suite has eighteen behavioural rubrics. The retail plugin ships two skills: retail-product-search, which builds a single-agent catalog search on Vector Search 2.0 and BigQuery, and retail-virtual-tryon, which uses Gemini image models and Veo 3.1 for composites and short videos.

Contrib includes twelve Python agents, among them a retail product-search service that reads from the collection the plugin provisions and can be deployed in five shapes. Plugin Recipes can be installed into coding assistants with the npx skills CLI. The project states that it is not an officially supported Google product and that the recipes are meant as demonstrations and starting points.

Developers building agents on Google Cloud can scaffold a starting agent with Agents CLI, then install a vertical skill into a coding assistant with a single command. Teams can reuse the retail plugin's setup, which provisions the BigQuery catalog and embeddings, instead of building that ingestion pipeline from scratch. Treat the recipes as starting points, since Google says they are not meant for production use.