ADK Agent Builder
Build production-ready agents with Google’s Agent Development Kit (ADK): scaffolding, tool wiring, orchestration patterns, testing, and optional deployment to Vertex AI Agent Engine.
Overview
-
Creates a minimal, production-oriented ADK scaffold (agent entrypoint, tool registry, config, and tests).
-
Supports single-agent ReAct-style workflows and multi-agent orchestration (Sequential/Parallel/Loop).
-
Produces a validation checklist suitable for CI (lint/tests/smoke prompts) and optional Agent Engine deployment verification.
Prerequisites
-
Python runtime compatible with your project (often Python 3.10+)
-
google-adk installed and importable
-
If deploying: access to a Google Cloud project with Vertex AI enabled and permissions to deploy Agent Engine runtimes
-
Secrets available via environment variables or a secret manager (never hardcoded)
Instructions
-
Confirm scope: local-only agent scaffold vs Vertex AI Agent Engine deployment.
-
Choose an architecture:
-
Single agent (ReAct) for adaptive tool-driven tasks
-
Multi-agent system (specialists + orchestrator) for complex, multi-step workflows
-
Define the tool surface (built-in ADK tools + any custom tools you need) and required credentials.
-
Scaffold the project:
-
src/agents/ , src/tools/ , tests/ , and a dependency file (pyproject.toml or requirements.txt )
-
Implement the minimum viable agent and a smoke test prompt; add regression tests for tool failures.
-
If deploying, produce an adk deploy ... command and a post-deploy validation checklist (AgentCard/task endpoints, permissions, logs).
Output
-
A repo-ready ADK scaffold (files and directories) plus starter agent code
-
Tool stubs and wiring points (where to add new tools safely)
-
A test + validation plan (unit tests and a minimal smoke prompt)
-
Optional: deployment commands and verification steps for Agent Engine
Error Handling
-
Dependency/runtime issues: provide pinned install commands and validate imports.
-
Auth/permission failures: identify the missing role/API and propose least-privilege fixes.
-
Tool failures/rate limits: add retries/backoff guidance and a regression test to prevent recurrence.
Examples
Example: Scaffold a single ReAct agent
-
Request: “Create an ADK agent that summarizes PRs and proposes test updates.”
-
Result: agent entrypoint + tool registry + a smoke test command for local verification.
Example: Multi-agent orchestrator
-
Request: “Build a supervisor + deployer + verifier team and deploy to Agent Engine.”
-
Result: orchestrator skeleton, per-agent responsibilities, and adk deploy ...
- post-deploy health checks.
Resources
-
Full detailed guide (kept for reference): ${CLAUDE_SKILL_DIR}/references/SKILL.full.md
-
Repo standards (source of truth):
-
000-docs/6767-a-SPEC-DR-STND-claude-code-plugins-standard.md
-
000-docs/6767-b-SPEC-DR-STND-claude-skills-standard.md
-
ADK / Agent Engine docs: https://cloud.google.com/vertex-ai/docs/agent-engine