Answers from the documentation; checking which model is on duty
Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
google/adk-recipes: a single source of truth for agent development - OpenSmartRoute
google/adk-recipes: a single source of truth for agent development
Google releases google/adk-recipes, a single repository that holds core agent patterns, plugin skills, and end‑to‑end deployable services, replacing the older adk‑samples.
Key points
The repo splits into core/, plugins/, and contrib/ folders for patterns, skills, and services.
Each recipe ships a manifest.yaml, AGENTS.md, and a Dockerfile for CI validation.
The retail product‑search skill uses Gemini Enterprise Vector Search 2.0 and BigQuery.
Horizon, a long‑horizon harness, is shipped under core/python/long‑horizon‑harness.
Why it matters: Enable developers to scaffold, test, and deploy agent solutions from a single source, cutting setup time and ensuring consistency.
By OpenSmartRoute editorial · written through the router by openai/gpt-oss-20b@ollama-gpt-oss-20b
Google released a new repository called google/adk-recipes. This repo acts as the single source of truth for agent development. It replaced the older google/adk-samples repository. Developers can now find core patterns, plugin skills, and deployable services in one place. The team curated these recipes to make agent creation easier and faster.
Previously, developers had to search multiple locations for code examples. Now, a single tree holds everything needed to build agents. This structure supports the Agents CLI, Agent Garden, and Gemini Enterprise app. Each of these tools reads from this central repository through its own protocol. The goal is to reduce friction when building new AI applications.
Repository structure: core, plugins, and contrib folders
The repository splits its content into three top-level folders. The first folder is named core. It holds reference patterns curated by the Agents CLI and DevRel teams. The second folder is named plugins. It contains installable vertical skills packaged as Plugin Recipes. These skills follow the Agent Plugins v1.0.0 spec. The third folder is named contrib. It holds end-to-end deployable services across five languages.
Every piece of code in these folders is designed for machine readability. This means scripts and configurations are easy for agents to parse. Humans can still read the README.md files for context. The AGENTS.md files tell coding assistants what a recipe is and how to use it. The manifest.yaml files provide a validated manifest for every recipe.
Standard recipe artifacts: manifest, AGENTS.md, Dockerfile, CI checks
Every recipe ships with specific validated artifacts. It includes a manifest.yaml file and an AGENTS.md file. Some recipes also include a Dockerfile and a runnability test. Continuous Integration enforces standards on every Pull Request. A suite of coding-assistant skills prepares the recipe before submission. These skills handle scaffolding, validation, and submission end-to-end.
The prepare-python-recipe skill helps with Python projects. The generate-manifest skill creates the necessary manifest files. The github-pr-review skill checks for quality issues. Other skills assist in the preparation process. This automation ensures that every recipe meets the required standards. Developers no longer need to manually verify every file.
Apigee hybrid 1.17.0 now lets AI agents use APIs inside your own cluster. The Model Context Protocol runs where your code lives without moving the runtime.
Long‑horizon harness Horizon and its evaluation suite
Building an agent that remembers across days is a hard task. The hard parts involve cross-session memory and sandboxing. The team built a reference harness called Horizon to collect these primitives. Horizon lives at core/python/long-horizon-harness in the repository. It serves as a production template for developers.
Developers can clone Horizon with a single command. They can then run make deploy to roll the backend to Cloud Run. The evaluation suite lives in the tests directory. It contains eighteen named behavioural rubrics. These rubrics cover memory recall, injection attempts, and guardrails. The rubrics execute with uv run pytest tests/eval.
A first chat turn confirms the memory layer is wired correctly. This harness helps coding assistants study before building harder agents. When a developer needs long-horizon capabilities, they use Horizon as a scaffold. The rest of the harness falls into place automatically.
Retail plugin skills: product‑search and virtual‑tryon
The second layer of the repository holds Plugin Recipes. The first vertical covered in this launch is retail. It is available under plugins/retail/. The Plugin Recipe format sits on top of the Agent Plugins v1.0.0 specification. This spec defines a plugin as a bundle of one or more skills. Each skill has a SKILL.md file and an EVAL.yaml file.
Two skills ship in this plugin today. The first skill is retail-product-search. It builds an ADK agent with a vector-search retrieval tool. The tool is backed by Vector Search 2.0 on the Gemini Enterprise Agent Platform. It also uses a BigQuery catalog for data storage. The skill handles the entire pipeline from ingestion to deployment.
The second skill is retail-virtual-tryon. It generates a photorealistic composite of a model wearing a product. It can also generate an eight-second studio catwalk video. The skill uses gemini-2.5-flash-image by default. It falls back to gemini-2.5-pro-image when quality matters more than speed. For video, it uses veo-3.1-generate-001 in Reference-to-Video mode.
Both skills install into any coding assistant with a single command. The command uses npx skills add google/adk-recipes. The --full-depth flag discovers the nested plugins/retail/skills structure. The install lands in the ~/.agents/skills directory. It is symlinked into every detected coding assistant on the local machine.
Why it matters: faster, consistent agent pipelines
This new structure enables faster and more consistent agent pipelines. Developers can scaffold a long-horizon memory agent from the terminal. They can install a retail vertical skill with one line of code. They can deploy a fully-wired production service to Cloud Run or Agent Garden. All of this comes from the same curated source tree.
Consistency reduces errors in production environments. The standardized artifacts ensure that every recipe follows the same rules. Coding assistants can open the repo and understand the structure immediately. This reduces the time needed to onboard new developers. Teams can ship features to customers much faster.
How OpenSmartRoute helps: routing requests through the new repo
A team that routes requests through OpenSmartRoute gains access to a curated catalogue of agent patterns. OpenSmartRoute scores every candidate on quality, cost, speed, and safety. The team sets the weights per request to optimize performance. Hard rules pin a request to specific models or regions. For example, text with personal data stays on an on-premises model.
The team can set up a new catalogue entry for google/adk-recipes. This entry competes on the next request. Nothing else changes in the app when a new model is added. OpenSmartRoute learns from outcomes, so a model that answers well gets more traffic. A model that fails gets less traffic. The savings ledger shows what each routed request cost next to the most expensive model.
Input guard spots prompt injection and personal data before a request leaves. The 'osr eval' tool measures routing accuracy on the team's own prompts. It can fail a build when the accuracy drops. The hosted platform keeps a models catalogue with prices and public rankings. This helps the team make data-driven decisions about model selection.
google/adk-recipes: a single source of truth for agent development
Google moved the ADK recipes into one main folder named google/adk-recipes. This folder replaced the old google/adk-samples location. The new tree acts as the single source of truth for all agent content. Every recipe inside must follow strict rules before it can be used.
The repository splits into three main top-level folders. The first folder is named core. It holds reference patterns curated by Google teams. The second folder is named plugins. It contains installable vertical skills. The third folder is named contrib. It stores end-to-end deployable services.
Every recipe inside these folders ships a specific file named manifest.yaml. This file acts as a validated manifest for the recipe. Each recipe also includes a file named AGENTS.md. This file tells coding assistants what the recipe is and how to use it. Some recipes also include a Dockerfile and a runnability test.
CI systems enforce standards on every pull request. A suite of coding-assistant skills prepares the recipes automatically. These skills include prepare-python-recipe and generate-manifest. They also include github-pr-review and others. Your agent can scaffold, validate, and submit a recipe end-to-end.
Every AGENTS.md file is written primarily for the AI coding assistant. The coding assistant opens the repo on your behalf. Every manifest.yaml file follows the same structure. Every SKILL.md file follows the same structure. README.md files are what humans still read.
The handbook glossary captures the dual framing of the content. A recipe is a runnable agent example. It can be found under core, contrib, or plugins. It is consumed by ADK developers and coding agents alike.
adk-recipes serves as the single source of truth for ADK content. It powers three Agent Platform surfaces that consume it. The first surface is Agents CLI. It reads the core folder directly. When a developer runs a specific command, Agents CLI can scaffold on demand. The command is uvx --from google-agents-cli agents-cli create my-agent.
The CLI launched earlier this year at Google Cloud Next. It exposes the full scaffold to deploy to evaluate to publish lifecycle. The base content it serves comes from our recipes. Horizon is an example pattern Agents CLI treats as a production template.
Agent Garden features a curated selection of recipes. The developer-facing path into the repository is a pull request to contrib. A recipe lands in GitHub and becomes part of the public ADK ecosystem. From that moment, it is discoverable by other developers. Coding assistants reading the repository also find it.
Appearing in the Agent Garden gallery is a separate editorial decision. The Agent Garden team curates a subset of recipes. They promote them to the public catalog on their own cadence. Once an agent is in the gallery, any developer in a Google Cloud project gets a one-click install.
Beyond these primary surfaces, Google Cloud infrastructure components appear as building blocks. Horizon uses Agent Platform Sandboxes for per-user sandboxing. It uses Memory Bank for durable memory. It uses Agent Platform Sessions for session state.
The retail product-search Plugin Recipe uses Vector Search 2.0. It runs on Gemini Enterprise Agent Platform. It pairs with BigQuery for data storage. The retail virtual-tryon Plugin Recipe uses Gemini image models. It creates still composites for users. It uses Veo 3.1 Reference-to-Video for animated outputs.
The recipes do not just live alongside Agent Platform. They document it in use for everyone. For developers working with coding assistants outside Google Cloud, Plugin Recipes install through the open npx skills CLI. This CLI is maintained by vercel-labs.
Building an agent that can handle a single conversation is a well-understood problem. Building one that holds its own across days of interaction is a materially different undertaking. It must remember what the user did yesterday. It must plan tasks for tomorrow.
The hard parts of such a long-horizon system are the same regardless of the agent's purpose. Cross-session memory is a major requirement. Per-user sandboxing of tool execution is another requirement. Secrets management is critical for security. Guardrails against runaway behaviour prevent errors. Orchestration primitives make multi-step tasks resumable.
Horizon lives at core/python/long-horizon-harness. It is the reference harness we built to collect these primitives. It puts them all in one place for developers.
The shortest path from a fresh clone to a running Horizon instance involves several steps. First, you run gcloud config set project YOUR_PROJECT_ID. Next, you run gcloud services enable aiplatform.googleapis.com. Then you run git clone --depth 1 --filter=blob:none --sparse https://github.com/google/adk-recipes.git. Finally, you run git sparse-checkout set core/python/long-horizon-harness.
By default, the environment backend is local. Sessions are in-memory. No cloud deploys are needed initially. You can run make deploy to roll the backend to Cloud Run. You use Terraform for the deployment. You can run make destroy to tear it back down.
Horizon's evaluation suite lives in tests/eval/evalsets. It contains eighteen named behavioural rubrics. These rubrics cover memory recall. They cover injection-based exfiltration attempts. They cover guardrail halting under adversarial input. They cover tool-selection discipline. They cover a journey-level plot-artifact evaluation.
The rubrics are executed with uv r and more consistent agent pipelines. Developers can scaffold a long-horizon memory agent from the terminal. They can install a retail vertical skill with one line of code. They can deploy a fully-wired production service to Cloud Run or Agent Garden. All of this comes from the same curated source tree.
Consistency reduces errors in production environments. The standardized artifacts ensure that every recipe follows the same rules. Coding assistants can open the repo and understand the structure immediately. This reduces the time needed to onboard new developers. Teams can ship features to customers much faster.
What to do
Read the AGENTS.md files in the google/adk-recipes repository. Check the manifest.yaml files to understand the structure. Clone the repository and run the Horizon example to test long-horizon capabilities. Install the retail-product-search skill in your coding assistant to see the plugin format in action.
Compare the new plugin format with your current agent development workflow. Identify which parts of your pipeline can benefit from the standardized artifacts. Consider using OpenSmartRoute to route traffic to these new recipes. Monitor the 'osr eval' results to ensure routing accuracy remains high.
