Google ADK Agent Garden — Enterprise Agent Inventory

See also (wiki): wiki/enterprise-agent-runtime-infrastructure.md · wiki/agentic-ai-governance.md · wiki/assistive-to-agentic-shift.md · wiki/workflow-redesign.md · wiki/agent-identity-permissions.md

Source credibility: HIGH — Google-authored reference implementations on ADK 2.0 from github.com/google/adk-samples (67 agents) and github.com/Google-Cloud-AI/agent-platform. Opened to all developers November 7, 2025, no GCP account required. These are canonical patterns, not community samples. No independent n= survey data. These case studies are vendor-published and represent selected wins with no control group and no independent verification. Cross-reference against: METR RCT (experienced developers 19% slower), CMU study (40.7% code complexity increase), Atlan 200-deployment analysis (median +159.8% ROI requires workflow redesign first).

Executive Summary

The Google ADK Agent Garden is a curated library of 67+ production-ready agent samples spanning financial services, healthcare, travel, customer service, business intelligence, and software development. Built on ADK 2.0 (LlmAgent, SequentialAgent, ParallelAgent, LoopAgent), the samples demonstrate the canonical enterprise agentic patterns: multi-agent orchestration, HITL gates via callbacks, Firestore-backed state persistence with repair/resume, LLM-as-Judge validation, and integration with GCP services (BigQuery, Vertex AI, Apigee). The architecture is Google Cloud-opinionated but framework-agnostic at the model layer. Note: there is no “Agent Core” product from Google — the managed hosting layer is “Agent Runtime” and the catalog layer is “Agent Registry.”

Key Data Points

ADK Framework Architecture

Core Classes

from google.adk.agents import (
    LlmAgent,         # Primary class — LLM + tools + callbacks + planner
    BaseAgent,        # Pydantic base; all agents inherit; defines run() async generator
    SequentialAgent,  # Runs sub-agents in order, passes state between them
    ParallelAgent,    # Runs sub-agents concurrently
    LoopAgent,        # Repeats until termination condition
)
from google.adk.tools.agent_tool import AgentTool   # Wraps an agent as a callable tool
from google.adk.planners import BuiltInPlanner       # ReAct-style with ThinkingConfig

Key Patterns

• Multi-agent composition: Sub-agents wrapped as AgentTool(sub_agent) and passed into parent’s tools=[]. Orchestrator LLM decides when to invoke each.
• State management: Session state dict. Sub-agents write to output_key; parent reads from state. Firestore backs this for persistence + repair/resume on failure.
• Callbacks: before/after_agent_callback, before_tool_callback — used for HITL gates, LLM-as-Judge validation, state initialization.
• Graph workflows (ADK 2.0): WorkflowAgent with edges=[(node_a, node_b)] mixes Python function nodes with LlmAgent nodes in one DAG.
• Tool types: FunctionTool (any Python fn), AgentTool, ApplicationIntegrationToolset (GCP no-code), MCP toolsets, built-in Google Search, BigQuery tools, code executor.

There is no “Agent Core” product by this name. The base is BaseAgent (Pydantic). Google has “Agent Runtime” (managed hosting) and “Agent Registry” (catalog) — not “Agent Core.”

Enterprise Agent Inventory

Financial Services

Healthcare / Insurance

Operations / Supply Chain

Security / Compliance

Software Engineering / DevOps

Marketing / Retail / Commerce

Research / Knowledge Work

Infrastructure Patterns

Portability to Non-GCP / Anthropic SDK

High Portability (1–4 days each)

These agents are prompt + Python function tools with no GCP infrastructure coupling:

• global-kyc-agent — Companies House + SEC EDGAR APIs are public REST. Port AgentTool composition to tool_use in Anthropic SDK directly.
• economic-research-agent — FRED, BLS, Census APIs are all public. No GCP dependency in core logic.
• llm-auditor — Pure reasoning + search. Replace built-in Google Search with any search tool.
• sdlc-task-planner / sdlc-technical-designer — Reasoning-only. Swap model + class definition.
• cyber-guardian-agent — BigQuery is the only GCP touch. Replace with any SQL store. The dynamic routing logic (conditional orchestration) maps directly to Claude’s tool-use loop.
• financial-advisor — Google Search only. Trivially portable.

Moderate Portability (1–2 weeks each)

Core logic clean; infrastructure bindings need abstraction:

• small-business-loan-agent — Replace Firestore → any KV store for state/resume. Replace GCS → S3/local. The HITL gate + LLM-as-Judge + repair/resume pattern is the valuable IP — architecture ports cleanly.
• medical-pre-authorization / claim-adjudication-agent — Replace Cloud Storage + Cloud Run triggers with equivalent. Core document processing logic is portable.
• invoice-processing — The dual-mode + 3-layer validation + ALF correction engine is sophisticated self-contained Python. Not GCP-coupled in core. Worth porting as-is.
• supply-chain — Replace BigQuery with Postgres/DuckDB for NL-to-SQL. WeatherNext has public equivalents. Core 5-specialist orchestration pattern is clean.

Lower Portability (deep GCP coupling)

• order-processing — depends on ApplicationIntegrationToolset (GCP-only no-code integration platform). No cross-cloud equivalent.
• ambient-expense-agent — Pub/Sub event trigger, IAP approval UI, Terraform GCP infra. Logic is portable; infra wrapper is not.
• data-engineering — Dataform (GCP-only).
• agent-observability-bq — BigQuery-specific telemetry.

ADK → Anthropic SDK Translation Map

Bottom line: The five highest-value ports for Fortune 500 enterprise use are global-kyc-agent, small-business-loan-agent, cyber-guardian-agent, supply-chain, and invoice-processing. All demonstrate HITL, LLM-as-Judge, parallel retrieval, and/or repair/resume — the patterns that matter most for enterprise deployment.

What This Means for Your Organization

The ADK Agent Garden solves a specific problem: enterprise teams that want to build production agents on Google Cloud no longer need to assemble orchestration, state management, HITL gates, and evaluation tooling from first principles. The 67 reference samples provide copy-and-adapt starting points for the highest-value enterprise domains (financial services, healthcare, supply chain, legal/compliance).

The patterns that matter most for regulated enterprise deployment are all demonstrated: Firestore-backed repair/resume for long-running workflows, before_tool_callback HITL gates before consequential actions, LLM-as-Judge validation at the output boundary, and parallel retrieval via ParallelAgent for latency-sensitive retrieval pipelines. These are not incidental features — they are the architecture requirements for any agentic system that needs to operate in a regulated environment with an audit trail.

The Agent Registry pattern (github.com/Google-Cloud-AI/agent-platform) addresses the governance problem that emerges when multiple teams deploy agents independently: centralized discovery, version control, and cross-team reuse. If your organization is past the pilot stage and has multiple agent deployments in flight, an agent registry is the next infrastructure investment to make.

Questions about Google ADK architecture for your use case? Reach out at brandon@brandonsneider.com.

Researched 2026-05-19. Source: github.com/google/adk-samples (67 agents), github.com/Google-Cloud-AI/agent-platform.

Brandon Sneider | brandon@brandonsneider.com May 2026