Google Releases adk-recipes for Standardized Agent Development
A unified repository structure for scaffolding, testing, and deploying production-ready AI agents.
1Unified Agent Source Tree
Google has restructured its ADK samples into the google/adk-recipes repository, creating a single source of truth for agent development. This new architecture organizes content into core, plugins, and contrib folders. It enables developers to scaffold, test, and deploy production services like Cloud Run or Gemini Enterprise apps from one location.
2Three-Layer Repository Architecture
The repository is divided into three distinct layers to serve different development needs. The core folder provides reference patterns for CLI teams, plugins contain installable vertical skills following the Agent Plugins v1.0.0 spec, and contrib hosts end-to-end deployable services. Every component is designed for machine readability to support automated agent workflows.
3Automated Agent-Friendly Manifests
Every recipe includes a validated manifest.yaml and an AGENTS.md file, specifically written for AI coding assistants. These files provide instructions on how to use the recipe, while CI pipelines enforce standards on every pull request. This setup allows coding assistants to scaffold, validate, and submit recipes without requiring manual human intervention.
4Agents CLI and Garden
The Agents CLI, launched at Google Cloud Next, uses core recipes to scaffold production templates. Meanwhile, the Agent Garden serves as a curated gallery for contrib recipes. Once a recipe is promoted to the gallery, developers can perform one-click installations, streamlining the transition from a GitHub pull request to a public catalog.
5Horizon Reference Harness
Horizon, located in core/python/long-horizon-harness, serves as a reference harness for building agents that operate over extended timeframes. It addresses critical challenges like cross-session memory, per-user sandboxing, and secrets management. By using Horizon as a scaffold, developers can ensure their agents handle multi-step tasks and maintain state across days of interaction.
6Rigorous Evaluation Rubrics
Horizon includes eighteen behavioral rubrics within its evaluation suite to test memory recall and guardrail performance. Developers execute these tests using pytest to ensure tool-selection discipline and prevent exfiltration attempts. This testing framework provides a baseline for quality, ensuring that agents remain stable and secure before they are deployed to production environments.
7Retail Plugin Vertical
The retail folder represents the first industry-specific vertical in the plugins layer. It adheres to the Agent Plugins v1.0.0 specification, allowing for consistent installation across various coding assistants. This vertical includes specialized skills for product search and virtual try-on, demonstrating how domain-specific logic can be packaged for easy developer adoption.
8Product Search Skill Pipeline
The retail-product-search skill automates the entire pipeline from data ingestion to deployment. It uses Vector Search 2.0 on the Gemini Enterprise Agent Platform and BigQuery to manage product catalogs. The skill handles embedding generation and vector collection building, allowing developers to deploy a functional retrieval agent with minimal manual configuration or setup.
9Virtual Try-On Capabilities
The virtual-tryon skill utilizes Gemini image models for photorealistic composites and Veo 3.1 for studio-quality catwalk animations. It includes a pre-flight product-cutout classifier and configurable safety levels to ensure editorial quality. These operational details are designed to be easily integrated into retail applications, supporting both still image generation and animated video outputs.
10Cross-Assistant Skill Installation
Developers can install skills using the npx skills CLI, which supports various coding assistants like Claude Code, Cursor, and Windsurf. The --full-depth flag allows the CLI to discover nested structures within the repository. Once installed, skills are symlinked locally, making them immediately available for use within the developer's preferred AI-assisted coding environment.
11Interactive Q-Mode Interviews
Upon installation, skills initiate a Q-Mode interview to configure the project based on specific requirements. The assistant asks for parameters like GCP project ID and region, then generates a design-spec.md file. This interactive process ensures that the agent is correctly provisioned and validated against quality gates before it begins active operations.
12Scaling Agent Ecosystems
The adk-recipes structure provides a scalable model for building complex, multi-step AI agents. By separating reference patterns, vertical plugins, and deployable services, Google enables developers to reuse components across different industries. This standardized approach reduces the complexity of agent lifecycle management, allowing teams to focus on building high-value, production-ready AI applications.