Adk Vertex Genai

ADK Vertex Gen AI

Operating Stance

Build Python Google Agent Development Kit (ADK) agents for Vertex AI Agent Engine and Gemini Enterprise. Treat Vertex AI as the production control plane and runtime target; do not use Google Developers Gemini Application Programming Interface (API) key patterns for this workflow.

Use this skill when work touches any of these areas:

• ADK LlmAgent, SequentialAgent, ParallelAgent, callbacks, tools, evaluation, or local ADK servers.
• Vertex AI Agent Engine source-based deployment, runtime service accounts, logs, or remote smoke tests.
• Google Gen AI Software Development Kit (SDK) calls through Vertex AI, especially Part.from_uri document understanding.
• Discovery Engine or Vertex AI Search retrieval tools, signed source links, and citation formatting.
• Gemini Enterprise registration of deployed ADK agents.

Workflow Map

flowchart TD
    A["Inspect repo and reference deployment"] --> B["Confirm Vertex AI-only config"]
    B --> C["Shape ADK package and app wrapper"]
    C --> D["Implement agents, workflow agents, tools, and callbacks"]
    D --> E["Validate Application Default Credentials and retrieval directly"]
    E --> F["Run local ADK server and agent checks"]
    F --> G["Deploy to Vertex AI Agent Engine from source"]
    G --> H["Read Agent Engine logs"]
    H --> I["Smoke-test remote Agent Engine"]
    I --> J["Register or prepare Gemini Enterprise registration"]
    J --> K["Verify Gemini Enterprise answer, sources, and latency"]

Core Rules

• Prefer source-based Vertex AI Agent Engine deployment for real repos. Use source_packages, entrypoint_module, entrypoint_object, generated requirements, and deployment metadata.
• Export root_agent from an importable package and wrap it with AdkApp(agent=root_agent) for Agent Engine.
• Separate local operator configuration from runtime configuration. Local files can include service account key paths; Agent Engine runtime environment variables must not include GOOGLE_APPLICATION_CREDENTIALS, GOOGLE_CLOUD_PROJECT, or local Gemini Enterprise registration variables.
• Use custom function tools for controlled retrieval, full-document inspection, signed links, and normalized citations. Prefer custom tools when built-in tool limitations or output shape matter.
• Use callbacks when behavior belongs around a model or tool boundary: normalize tool output, collect source state, add final citations, enforce markdown shape, or hide intermediate workflow output.
• Read logs and run remote queries before declaring a deployment complete.

Vertex AI Gen AI SDK Rules

The Google Gen AI SDK must run in Vertex AI mode for this workflow.

• Preferred explicit client:

from google import genai
from google.genai import types

client = genai.Client(
    vertexai=True,
    project=project_id,
    location="global",
    http_options=types.HttpOptions(api_version="v1"),
)

• Environment-variable mode is acceptable only if it is clearly Vertex AI mode: GOOGLE_GENAI_USE_VERTEXAI=True, GOOGLE_CLOUD_PROJECT, and GOOGLE_CLOUD_LOCATION.
• Do not use GOOGLE_API_KEY, Google AI Studio key flows, or non-Vertex Gemini Developer API clients for Agent Engine production agents.
• Use types.Part.from_uri(file_uri="gs://...", mime_type="application/pdf") for Cloud Storage document understanding when the runtime service account has object read access.
• Use global for Gemini 3 preview model endpoints unless a current official model page says otherwise. Note in the plan that Gemini 3 models are preview if the official model page says preview.
• Keep model settings configurable: primary agent model, document-understanding model, location, output token limits, media resolution, and thinking level.

Agent Design Rules

• Keep the root agent as the front door when future domain expansion is likely. Its job is greeting, capability explanation, high-level guardrails, and delegation.
• Put domain retrieval and policy answering in specialized sub-agents.
• For deep-research behavior, use a sequential-into-parallel-into-synthesis pattern:
• planner decomposes the question,
• parallel research agents call retrieval and document tools,
• synthesizer produces the only final user-visible answer,
• callback or workflow wrapper suppresses intermediate text while preserving tool events and state deltas.
• If a workflow wrapper hides intermediate emissions, never drop function-call or function-response events. Those events are needed for ADK state and tool execution.
• All grounded-answer agents must abstain when retrieval or document evidence is insufficient.
• Final answers from document agents should cite distinct sources once, preserve markdown, and use valid markdown links.

Deployment Rules

• Inspect any reference repo the user names before changing deployment commands. Mirror the working make deploy and make register-gemini-enterprise shape when the team expects that interface.
• Change into repo root before packaging.
• Put generated requirements inside the uploaded source package and pass requirements_file relative to the uploaded source root.
• Pin or constrain core dependencies enough for repeatable Agent Engine builds: google-cloud-aiplatform[agent_engines,adk], google-adk, google-genai, google-cloud-discoveryengine, google-cloud-storage, and any framework-specific packages.
• Use a dedicated runtime service account. Grant minimum practical roles for the actual tools: Agent Engine runtime invocation, Discovery Engine or Vertex AI Search read/search, Cloud Storage object read, and Service Account Token Creator only if signing URLs through Identity and Access Management (IAM) credentials.
• Write deployment_metadata.json after successful create or update.

Verification Rules

• Validate Application Default Credentials (ADC) locally before agent tests.
• Test retrieval and full-document tools directly before prompt debugging.
• Start a local ADK server and test at least one greeting plus one retrieval-backed question.
• After deploy, create a remote session and stream-query the deployed Agent Engine resource.
• Benchmark latency and quality when changing model families, thinking levels, document reads, or parallel lane count.
• For Gemini Enterprise, test the rendered output, not just the raw Agent Engine stream. Broken markdown can be invisible locally but ugly in Enterprise.

Reference Files

• Read references/checklists.md before implementation, deployment, and handoff.
• Read references/common-failures.md when local, deployment, retrieval, document, signed-link, or Gemini Enterprise behavior breaks.
• Read references/vertex-genai-patterns.md for Vertex AI Gen AI SDK code patterns, document-understanding tools, callbacks, and deep-research workflow guidance.

Official References

• ADK overview: https://adk.dev/get-started/about/
• ADK agents: https://adk.dev/agents/
• ADK sequential workflow agents: https://adk.dev/agents/workflow-agents/sequential-agents/
• ADK parallel workflow agents: https://adk.dev/agents/workflow-agents/parallel-agents/
• ADK callbacks: https://adk.dev/callbacks/
• ADK tool limitations: https://adk.dev/tools/limitations/
• ADK evaluation criteria: https://adk.dev/evaluate/criteria/
• Vertex AI Agent Engine deployment: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/agent-engine/deploy
• Gemini Enterprise ADK registration: https://docs.cloud.google.com/gemini/enterprise/docs/register-and-manage-an-adk-agent
• Agent Starter Pack deployment: https://googlecloudplatform.github.io/agent-starter-pack/guide/deployment.html
• Agent Starter Pack Gemini Enterprise registration command: https://googlecloudplatform.github.io/agent-starter-pack/cli/registergeminienterprise.html
• Vertex AI Gen AI SDK overview: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/sdks/overview
• Vertex AI document understanding with Part.from_uri: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/multimodal/document-understanding
• Vertex AI model versions: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions
• Vertex AI deployments and endpoints: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/locations

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: andrey-learning-machines
• Source: andrey-learning-machines/swe-harness
• License: MIT

Install and usage instructions live in the source repository linked above.