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

1. Confirm scope: local-only agent scaffold vs Vertex AI Agent Engine deployment.
2. Choose an architecture:

• Single agent (ReAct) for adaptive tool-driven tasks
• Multi-agent system (specialists + orchestrator) for complex, multi-step workflows

1. Define the tool surface (built-in ADK tools + any custom tools you need) and required credentials.
2. Scaffold the project:

• src/agents/, src/tools/, tests/, and a dependency file (pyproject.toml or requirements.txt)

1. Implement the minimum viable agent and a smoke test prompt; add regression tests for tool failures.
2. 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

• Implementation patterns (scaffolds, tool wiring, orchestration): ${CLAUDE_SKILL_DIR}/references/implementation.md
• Worked examples (single-agent + multi-agent): ${CLAUDE_SKILL_DIR}/references/examples.md
• Error-handling and recovery patterns: ${CLAUDE_SKILL_DIR}/references/errors.md
• Product / architecture context: ${CLAUDE_SKILL_DIR}/PRD.md, ${CLAUDE_SKILL_DIR}/ARD.md
• ADK / Agent Engine docs: