ADK Agent Recipes
Start ADK agent projects from small, runnable reference implementations.
The workflows show unusually concrete static controls over privilege and effects: the model receives no GitHub write token, PR-head code is not checked out, workspace agent configuration is removed, app-token permissions are narrowed explicitly, and model decisions are validated against real issue numbers and label names. Credentials flow through WIF, secrets, and scoped environment variables, telemetry is disabled, and comments explain what data reaches the model and where writes occur. Deductions apply because dependency security is incomplete: Actions are commit-pinned, but the Antigravity CLI is installed as the latest version through a remote shell script and the shown Python runtime dependencies are unpinned. Label edits are manually reversible but lack automatic transactional rollback. Ambiguous single-issue classification is deferred to a human, yet validated classifications can still create automatic external writes. The license, verified organization provenance, and explicit disclaimer provide strong attribution.
The README, configuration, tests, and workflow commentary are broadly consistent. Guards and actionable messages cover empty or oversized diffs, argument-size limits, invalid model output, temporary API failures, and parse failures, with bounded retries where appropriate. Deductions apply because the supplied PR workflow is truncated, so its full validation and posting path cannot be confirmed statically, and recipe-specific operation is delegated to README files not supplied here. Dependency availability is addressed through Python requirements, development dependencies, a PyPI index, a minimum uv version, and retries, but reliance on a dynamically fetched CLI plus GitHub, Google Cloud, and ADK services reduces control.
The README identifies learners and project starters and names customer-service, research, and industry workflows, while separating canonical core patterns from broader community recipes. The workflows define precise boundaries for model tools, untrusted input, valid labels, in-batch issue numbers, diff size, and human fallback for uncertain classifications. Environment coverage includes several ADK language SDKs, Python constraints, and CI-specific safeguards. The deduction is that recipe-level variables, platform differences, and deployment requirements are absent from the supplied evidence, while some automation is tightly coupled to GitHub Actions, Google Cloud WIF, and Antigravity CLI.
The repository has clear core, contrib, skills, documentation, and self-contained recipe areas, with straightforward entry and contribution paths. Tests enforce project naming, directory placement, Python-version references, and pytest discovery, providing strong evidence of stable conventions. The complete Apache-2.0 license matches the README, and an issue path and official organization provenance are supplied. Deductions apply because top-level installation mainly redirects readers to external ADK and recipe-specific guides, no top-level FAQ is shown, and no changelog, release policy, or repository compatibility history is provided. The agents-cli team and Google organization indicate an update path, but the project expressly disclaims official support, limiting maintenance responsibility.
The recipes are framed as small, runnable, forkable starting points. The automation produces structured JSON, filters labels precisely, records summaries, and falls back visibly to human handling, making its outputs statically usable. The collection and automation plausibly add value over starting from scratch or handling all triage and review manually. Deductions apply because no executed outcomes, user-impact evidence, or resource benchmarks are supplied; the workflows acknowledge model calls, external services, token pressure, and diff truncation, while cloud and dynamically installed CLI dependencies add operational cost.
Most important safety and reliability claims map directly to adjacent permissions, YAML, shell guards, and unit tests. Design reasons, observed limitations, factual constraints, and policy choices are generally distinguished clearly. The README, pyproject, workflows, and tests corroborate repository organization, naming rules, test scope, and disclaimers. The deduction is that only a subset of the repository is supplied: actual recipe implementations, full documentation, the lockfile, the end of the PR workflow, and CI records are missing, so cross-source corroboration is good but incomplete.
- This is a low-confidence static assessment; no recipe, test, workflow, or model-output quality was executed or independently verified.
- Do not treat the samples as production-ready by default: the README explicitly describes them as demonstrations and starting points without official support.
- Antigravity CLI is fetched as the latest release through a remote installer without a reproducible version pin; pin and inspect the artifact before deployment.
- Individual recipes may have distinct network access, credentials, data flows, and external effects that the supplied evidence does not cover; review each recipe separately before adoption.
- The supplied PR-review workflow is truncated, so the complete final response-validation and comment-posting logic cannot be confirmed.
What does this agent do, and when should you use it?
This repository is a public collection of Agent Development Kit recipes rather than one deployable agent product. Its core/ directory contains canonical patterns curated by the agents-cli team, while contrib/ contains broader community examples for particular use cases and industry workflows. Documented pattern areas include OAuth flows, session memory, guardrails, and RAG, with each recipe acting as a self-contained example and carrying its own setup and run instructions. Adopters must install ADK first; the repository points to Python, TypeScript, Go, Java, and Kotlin SDKs. The collection is best treated as learning material or a project foundation because its disclaimer characterizes the recipes as demonstrations and starting points, not production-ready or officially supported Google products.
A developer selects a recipe under core/ or contrib/, follows that recipe's README.md, and runs the self-contained example. core/ demonstrates focused patterns maintained by the agents-cli team, including OAuth flows, session memory, guardrails, and RAG patterns. contrib/ supplies examples aimed at specific use cases or industry workflows. Developers can fork a recipe as the foundation of a new project or inspect the collection to learn ADK implementation patterns. Repository-maintenance resources include the Recipe Checklist, Recipe Handbook, and helpers under .agents/skills/ for tasks such as recipe scaffolding, manifest generation, and pyproject alignment; those helpers are distinct from user-facing vertical skills.
- A developer new to ADK who wants to begin with a small runnable example instead of an empty project.
- An engineer implementing OAuth, session memory, guardrails, or RAG who needs a focused reference pattern.
- A team prototyping a customer-service bot or research agent that wants to fork an existing foundation.
- A builder targeting a particular industry workflow who wants to inspect self-contained community recipes in contrib/.
- A contributor preparing a new recipe who needs the Recipe Checklist and Recipe Handbook for repository requirements.
What are this agent's strengths and limitations?
- The core/ and contrib/ split distinguishes curated canonical patterns from broader community use-case examples.
- Recipes are described as small, runnable, and self-contained, making them practical foundations for ADK experimentation.
- The collection names concrete engineering patterns such as OAuth, session memory, guardrails, and RAG.
- ADK SDKs are linked for Python, TypeScript, Go, Java, and Kotlin, supporting several implementation languages.
- Contribution resources include a concise checklist, a detailed handbook, and helpers for scaffolding, manifests, and pyproject alignment.
- This is not one cohesive product; installation, credentials, inputs, outputs, and execution differ by recipe.
- The top-level material supplies no universal install command, first-run command, API contract, or deployment path.
- Every recipe is built around the ADK ecosystem, so moving to another agent framework would require adaptation or reimplementation.
- contrib/ is community-driven, but the supplied material establishes no common maturity, testing, or maintenance guarantee for those recipes.
- The repository explicitly positions the recipes as demonstrations and starting points rather than production use, and it is not an officially supported Google product.
How do you install or deploy this agent?
The repository-level README only instructs users to install Agent Development Kit through the ADK Get Started guide and identifies Python, TypeScript, Go, Java, and Kotlin SDKs. It does not provide one copyable installation command or define a common language version, package manager, or credential set. Select a recipe in core/ or contrib/ and use the setup instructions in that recipe's README.md. Because no individual recipe documentation is included in the supplied material, exact dependency commands, environment variables, and credentials cannot be stated reliably.
How do you use this agent?
Choose a canonical OAuth, session-memory, guardrail, or RAG pattern from core/, or a use-case and industry-oriented example from contrib/. Open the selected recipe's README.md, complete its setup, and use its documented run procedure; alternatively, fork the recipe as a starting point for a new project. The supplied material documents no repository-wide CLI, API, deployment command, or first invocation, so there is no accurate universal command to provide. Contributors should begin with docs/recipe-checklist.md and use docs/recipe-handbook/README.md when they need the fuller tooling and process reference.