Google Updates Agent Development Kit with Unified Recipe Repository
Google has restructured its adk-recipes repository to organize agent patterns, vertical plugins, and deployable services into a single source of truth.
The three-layer architecture of the updated google/adk-recipes repository.
- Google has reorganized its ADK repository into three distinct layers: core, plugins, and contrib, providing a structured approach for AI agent development.
- The repository serves as a unified source for scaffolding agents, installing vertical skills, and deploying services to Google Cloud environments.
- Each recipe now includes validated manifests and automated runnability tests, specifically designed to be read and utilized by AI coding assistants.
Overview of the Restructured ADK Recipes
Google has introduced a significant restructuring of its google/adk-recipes repository, formerly known as google/adk-samples. This update aims to provide a centralized source of truth for developers building production-ready AI agents using the Agent Development Kit (ADK). The new organization categorizes content into three distinct layers, each tailored to specific developer and machine-based workflows.
By unifying these resources, the repository seeks to simplify the lifecycle of agent development, covering everything from initial scaffolding and skill installation to deployment across Google Cloud platforms, such as Agent Engine, Cloud Run, and the Gemini Enterprise app.
The Three-Layer Architecture
The repository content is now partitioned into three top-level directories, each serving a unique function:
- core/: This folder contains canonical patterns curated by the Agents CLI and Developer Relations teams. It focuses on fundamental building blocks for agents, such as OAuth flows, session memory management, guardrails, and Retrieval-Augmented Generation (RAG) patterns. It is designed to serve as a library of reference implementations.
- plugins/: This section holds installable vertical skills that conform to the Agent Plugins v1.0.0 specification. These plugins are intended to be integrated into coding assistants and development environments to provide specialized capabilities, such as retail-specific product search or virtual try-on tools.
- contrib/: This area contains end-to-end deployable services across various programming languages. Recipes here are community-contributed or vertical-specific, providing templates for developers to deploy functional agents into production environments.
Integration with AI Coding Assistants
A key focus of this update is the machine readability of the repository. Rather than prioritizing standard documentation exclusively for humans, the project team has optimized the directory structure and metadata for AI coding assistants. Each recipe is equipped with a manifest.yaml file and an AGENTS.md document, which provide structured instructions for assistants to understand and utilize the code.
Furthermore, the repository includes built-in coding-assistant skills in the .agents/skills/ directory, such as prepare-python-recipe and generate-manifest. These tools allow agents to automate the scaffolding, validation, and submission of new recipes, streamlining the development process. For developers using external tools beyond Google's ecosystem, such as Claude Code, Cursor, or Zed, the plugin recipes are designed to be compatible with the npx skills CLI maintained by vercel-labs.
Deployment and Operational Patterns
The updated repository serves as the backbone for several Google Cloud Agent Platform surfaces. The Agents CLI utilizes the core/ layer to scaffold new projects, while the Agent Garden gallery features curated recipes sourced from contrib/. These components are designed to work together, allowing developers to move from a local reference pattern to a deployed production service using standard commands like make dev-local and make deploy.
For instance, the "Horizon" reference harness, located in core/python/long-horizon-harness, demonstrates how to implement complex agent behaviors like cross-session memory and per-user sandboxing. By providing a standardized template for these difficult architectural challenges, the repository aims to help developers reduce the overhead of building resilient, long-horizon agents. It is important to note that these recipes are intended as demonstration material and starting points, and are not officially supported as enterprise products.
Enjoyed this?
Get more posts like this delivered to your inbox.