ADK Agent Builder Skill

Overview

The ADK Agent Builder accelerates AI agent development using Google's Agent Development Kit (ADK) framework integrated with Claude API. This skill provides production-ready scaffolding for:

• React-pattern agents (Reasoning + Acting loops)
• Multi-agent systems (Orchestrated agent teams)
• Tool-augmented agents (LinkedIn, Apollo, Clearbit, etc.)
• Workflow automation (Deterministic multi-step processes)
• State management (Context persistence across iterations)
• Testing frameworks (Unit, integration, E2E)

Installation

# Install Google ADK SDK
pip install google-adk anthropic>=0.18.0

# Install this plugin
/plugin install jeremy-google-adk@jeremylongshore

# Verify installation
adk-agent --version

Quick Start (5 Minutes)

Create Your First Agent

# Generate a LinkedIn intelligence agent
adk-agent create \
  --name linkedin-intelligence \
  --pattern react \
  --llm claude-3-5-sonnet \
  --tools linkedin-scraper,profile-analyzer \
  --output ./my-agent/

# Navigate and run
cd my-agent/
pip install -r requirements.txt
python linkedin_intelligence.py

Generated Structure

my-agent/
├── linkedin_intelligence.py    # Main agent implementation
├── tools/                      # Tool implementations
│   ├── linkedin_scraper.py
│   └── profile_analyzer.py
├── tests/                      # Test suite
│   └── test_agent.py
├── config/                     # Configuration
│   └── agent_config.yaml
├── requirements.txt            # Dependencies
├── Dockerfile                  # Container ready
└── README.md                   # Documentation

Agent Patterns

1. React Pattern (Single Agent)

Best for: Autonomous agents that think and act iteratively

from google_adk import ReactAgent, ToolRegistry
from anthropic import Anthropic
import os

class LinkedInIntelligenceAgent(ReactAgent):
    """
    Autonomous agent that researches LinkedIn profiles
    Uses React pattern: Thought -> Action -> Observation loop
    """

    def __init__(self):
        super().__init__()
        self.claude = Anthropic(api_key=os.getenv('CLAUDE_API_KEY'))
        self.model = "claude-3-5-sonnet-20241022"
        self.max_iterations = 10

        # Register tools
        self.tools = ToolRegistry([
            LinkedInScraperTool(),
            ProfileAnalyzerTool(),
            CompanyResearchTool(),
        ])

    async def run(self, profile_url: str) -> dict:
        """Execute React loop"""
        context = self.create_context(goal=f"Analyze {profile_url}")

        while not self.is_complete(context) and context.iterations < self.max_iterations:
            # Think: What should I do next?
            thought = await self.think(context)

            # Act: Execute the decided action
            observation = await self.act(thought)

            # Update: Store observation
            context.add_observation(observation)
            context.iterations += 1

        return self.synthesize_results(context)

    async def think(self, context):
        """Use Claude to reason about next action"""
        prompt = f"""You are analyzing a LinkedIn profile.

Goal: {context.goal}

Observations so far:
{self._format_observations(context.observations)}

Available tools: {', '.join(self.tools.list_tools())}

What should you do next? Respond with:
THOUGHT: [Your reasoning]
ACTION: [Tool to use]
PARAMS: [Tool parameters as JSON]
"""

        response = await self.claude.messages.create(
            model=self.model,
            max_tokens=1024,
            temperature=0.3,
            messages=[{"role": "user", "content": prompt}]
        )

        return self._parse_thought(response.content[0].text)

2. Multi-Agent System

Best for: Complex workflows requiring specialization

from google_adk import MultiAgentOrchestrator, Agent

class SDRAgentTeam:
    """
    Orchestrated team of specialized agents for sales development
    """

    def __init__(self):
        self.orchestrator = MultiAgentOrchestrator()

        # Initialize specialized agents
        self.research_agent = ResearchAgent()
        self.qualification_agent = QualificationAgent()
        self.outreach_agent = OutreachAgent()
        self.followup_agent = FollowUpAgent()

        # Define workflow
        self.workflow = [
            ("research", self.research_agent),
            ("qualify", self.qualification_agent),
            ("outreach", self.outreach_agent),
            ("followup", self.followup_agent),
        ]

    async def process_leads(self, lead_list: list) -> dict:
        """Process leads through the SDR pipeline"""
        results = {
            "researched": [],
            "qualified": [],
            "contacted": [],
            "meetings_booked": []
        }

        for lead in lead_list:
            # Research phase
            research = await self.research_agent.run(lead)
            results["researched"].append(research)

            # Qualification phase
            if research["score"] > 0.7:
                qualified = await self.qualification_agent.run(research)
                results["qualified"].append(qualified)

                # Outreach phase
                if qualified["is_icp"]:
                    outreach = await self.outreach_agent.run(qualified)
                    results["contacted"].append(outreach)

                    # Follow-up phase
                    if outreach["response_received"]:
                        followup = await self.followup_agent.run(outreach)
                        if followup["meeting_booked"]:
                            results["meetings_booked"].append(followup)

        return results

3. Workflow Pattern

Best for: Deterministic, repeatable processes

from google_adk import Workflow, Step

class DataEnrichmentWorkflow(Workflow):
    """
    Deterministic workflow for lead enrichment
    """

    def __init__(self):
        super().__init__()

        # Define workflow steps
        self.add_step(Step("validate", self.validate_input))
        self.add_step(Step("enrich_company", self.enrich_company_data))
        self.add_step(Step("enrich_person", self.enrich_person_data))
        self.add_step(Step("score", self.calculate_score))
        self.add_step(Step("route", self.route_to_team))

    async def validate_input(self, data):
        """Validate and clean input data"""
        # Implementation
        pass

    async def enrich_company_data(self, data):
        """Add company information from Clearbit/Apollo"""
        # Implementation
        pass

    async def enrich_person_data(self, data):
        """Add person information from LinkedIn/ContactOut"""
        # Implementation
        pass

    async def calculate_score(self, data):
        """Calculate lead score based on ICP criteria"""
        # Implementation
        pass

    async def route_to_team(self, data):
        """Route to appropriate sales team"""
        # Implementation
        pass

Tool Development

Creating Custom Tools

from google_adk import Tool, ToolResult

class LinkedInScraperTool(Tool):
    """
    Tool for scraping LinkedIn profiles
    """

    name = "linkedin_scraper"
    description = "Scrapes public LinkedIn profile data"

    def __init__(self):
        super().__init__()
        self.rate_limit = 60  # requests per minute
        self.cache_ttl = 3600  # 1 hour cache

    @property
    def input_schema(self):
        return {
            "type": "object",
            "properties": {
                "profile_url": {
                    "type": "string",
                    "description": "LinkedIn profile URL"
                },
                "include_activity": {
                    "type": "boolean",
                    "description": "Include recent activity",
                    "default": False
                }
            },
            "required": ["profile_url"]
        }

    async def call(self, profile_url: str, include_activity: bool = False) -> ToolResult:
        """Execute the tool"""
        try:
            # Check cache first
            cached = self.get_from_cache(profile_url)
            if cached:
                return ToolResult(success=True, data=cached)

            # Scrape profile
            profile_data = await self._scrape_profile(profile_url)

            if include_activity:
                activity = await self._scrape_activity(profile_url)
                profile_data["recent_activity"] = activity

            # Cache result
            self.cache_result(profile_url, profile_data)

            return ToolResult(
                success=True,
                data=profile_data,
                metadata={"source": "linkedin", "cached": False}
            )

        except Exception as e:
            return ToolResult(
                success=False,
                error=str(e),
                metadata={"tool": self.name}
            )

    async def _scrape_profile(self, url: str) -> dict:
        """Actual scraping logic"""
        # Implementation using Selenium/Playwright
        pass

Built-in Tools Library

# Available tools out of the box
from google_adk.tools import (
    # Web & Search
    WebSearchTool,           # Google search
    WebScraperTool,         # Generic web scraping

    # Sales & Marketing
    LinkedInTool,           # LinkedIn profiles/companies
    ApolloTool,            # Apollo.io enrichment
    ClearbitTool,          # Clearbit enrichment
    HunterTool,            # Email finder

    # Communication
    EmailTool,             # Send emails via SMTP/SendGrid
    SlackTool,             # Slack messaging
    CalendlyTool,          # Schedule meetings

    # Data & Analytics
    GoogleSheetsTool,      # Read/write sheets
    BigQueryTool,          # Query BigQuery

    # AI & Processing
    ClaudeTool,            # Claude API wrapper
    VertexAITool,          # Vertex AI integration

    # Utilities
    CalculatorTool,        # Math operations
    DateTimeTool,          # Date/time operations
    JSONTool,              # JSON parsing/manipulation
)

Claude Integration

Basic Claude Setup

from anthropic import Anthropic
from google_adk import ClaudeIntegration

class ClaudePoweredAgent:
    def __init__(self):
        self.claude = ClaudeIntegration(
            api_key=os.getenv('CLAUDE_API_KEY'),
            model="claude-3-5-sonnet-20241022",
            max_tokens=4096,
            temperature=0.3
        )

    async def process(self, input_text: str) -> str:
        """Process input using Claude"""
        response = await self.claude.complete(
            prompt=self.build_prompt(input_text),
            system="You are a helpful AI agent."
        )
        return response

Tool Calling with Claude

# Define tools for Claude
tools = [
    {
        "name": "search_linkedin",
        "description": "Search LinkedIn for profiles",
        "input_schema": {
            "type": "object",
            "properties": {
                "query": {"type": "string"},
                "filters": {"type": "object"}
            },
            "required": ["query"]
        }
    },
    {
        "name": "analyze_profile",
        "description": "Analyze a LinkedIn profile",
        "input_schema": {
            "type": "object",
            "properties": {
                "profile_url": {"type": "string"}
            },
            "required": ["profile_url"]
        }
    }
]

# Use tools in conversation
response = await self.claude.messages.create(
    model="claude-3-5-sonnet-20241022",
    max_tokens=4096,
    tools=tools,
    messages=[
        {
            "role": "user",
            "content": "Find and analyze the LinkedIn profile of the CEO of OpenAI"
        }
    ]
)

# Process tool calls
if response.stop_reason == "tool_use":
    for tool_use in response.content:
        if tool_use.type == "tool_use":
            result = await self.execute_tool(
                tool_use.name,
                tool_use.input
            )
            # Continue conversation with result

Configuration

Agent Configuration (agent_config.yaml)

agent:
  name: linkedin-intelligence-agent
  version: 1.0.0
  description: Analyzes LinkedIn profiles for sales intelligence

llm:
  provider: anthropic
  model: claude-3-5-sonnet-20241022
  max_tokens: 4096
  temperature: 0.3

tools:
  - name: linkedin_scraper
    enabled: true
    rate_limit: 60
    cache_ttl: 3600

  - name: clearbit_enrichment
    enabled: true
    api_key: ${CLEARBIT_API_KEY}

  - name: email_finder
    enabled: true
    provider: hunter

behavior:
  max_iterations: 10
  timeout_seconds: 300
  retry_on_error: true
  max_retries: 3

observability:
  logging:
    level: INFO
    format: json
    destination: stdout

  metrics:
    enabled: true
    provider: prometheus
    port: 9090

  tracing:
    enabled: true
    provider: opentelemetry
    endpoint: http://localhost:4317

Environment Variables

# Claude API
CLAUDE_API_KEY=sk-ant-your-key-here
CLAUDE_MODEL=claude-3-5-sonnet-20241022

# Google Cloud
GOOGLE_PROJECT_ID=your-project-id
GOOGLE_APPLICATION_CREDENTIALS=/path/to/credentials.json

# Third-party APIs
CLEARBIT_API_KEY=your-clearbit-key
APOLLO_API_KEY=your-apollo-key
HUNTER_API_KEY=your-hunter-key

# Observability
ENABLE_METRICS=true
METRICS_PORT=9090
LOG_LEVEL=INFO

Testing

Unit Tests

import pytest
from unittest.mock import Mock, patch
from linkedin_intelligence import LinkedInIntelligenceAgent

@pytest.fixture
def agent():
    """Create test agent instance"""
    return LinkedInIntelligenceAgent()

@pytest.mark.asyncio
async def test_agent_initialization(agent):
    """Test agent initializes correctly"""
    assert agent.model == "claude-3-5-sonnet-20241022"
    assert agent.max_iterations == 10
    assert len(agent.tools.list_tools()) > 0

@pytest.mark.asyncio
async def test_think_generates_valid_thought(agent):
    """Test thinking process generates valid action"""
    context = agent.create_context(goal="Test goal")

    with patch.object(agent.claude, 'messages') as mock_claude:
        mock_claude.create.return_value.content = [
            Mock(text="THOUGHT: Need to scrape\nACTION: linkedin_scraper\nPARAMS: {}")
        ]

        thought = await agent.think(context)

        assert thought["action"] == "linkedin_scraper"
        assert "thought" in thought
        assert "params" in thought

@pytest.mark.asyncio
async def test_complete_workflow(agent):
    """Test complete agent workflow"""
    profile_url = "https://linkedin.com/in/test"

    with patch.object(agent, 'tools') as mock_tools:
        mock_tools.execute.return_value = {
            "name": "Test User",
            "title": "CEO",
            "company": "Test Corp"
        }

        result = await agent.run(profile_url)

        assert result["success"] == True
        assert "observations" in result
        assert result["iterations"] > 0

Integration Tests

@pytest.mark.integration
class TestMultiAgentSystem:
    @pytest.mark.asyncio
    async def test_sdr_team_processes_leads(self):
        """Test SDR team processes leads correctly"""
        team = SDRAgentTeam()

        test_leads = [
            {"name": "John Doe", "company": "Tech Corp"},
            {"name": "Jane Smith", "company": "Sales Inc"}
        ]

        results = await team.process_leads(test_leads)

        assert len(results["researched"]) == 2
        assert len(results["qualified"]) > 0
        assert "meetings_booked" in results

    @pytest.mark.asyncio
    async def test_agent_coordination(self):
        """Test agents coordinate properly"""
        team = SDRAgentTeam()

        # Test research agent passes data to qualification
        research_output = await team.research_agent.run({"name": "Test"})
        qualification_input = await team.qualification_agent.validate_input(research_output)

        assert qualification_input is not None

End-to-End Tests

@pytest.mark.e2e
@pytest.mark.skipif(not os.getenv('CLAUDE_API_KEY'), reason="No API key")
class TestRealAgentExecution:
    @pytest.mark.asyncio
    async def test_real_linkedin_analysis(self):
        """Test real LinkedIn profile analysis"""
        agent = LinkedInIntelligenceAgent()

        # Use a public profile for testing
        result = await agent.run("https://linkedin.com/in/jeffweiner08")

        assert result["success"] == True
        assert "name" in result["data"]
        assert "company" in result["data"]
        assert result["data"]["name"] == "Jeff Weiner"

Deployment

Container Deployment

# Multi-stage build for agent
FROM python:3.11-slim AS builder

WORKDIR /app
COPY requirements.txt .
RUN pip install --no-cache-dir --user -r requirements.txt

FROM python:3.11-slim
RUN useradd -m -u 1000 agent

WORKDIR /app
COPY --from=builder /root/.local /home/agent/.local
COPY --chown=agent:agent . .

USER agent
ENV PATH=/home/agent/.local/bin:$PATH

EXPOSE 8080
HEALTHCHECK --interval=30s --timeout=3s \
  CMD curl -f http://localhost:8080/health || exit 1

CMD ["python", "-m", "agent.main"]

Cloud Run Deployment

# Build and push container
docker build -t gcr.io/my-project/linkedin-agent:latest .
docker push gcr.io/my-project/linkedin-agent:latest

# Deploy to Cloud Run
gcloud run deploy linkedin-agent \
  --image gcr.io/my-project/linkedin-agent:latest \
  --platform managed \
  --region us-central1 \
  --allow-unauthenticated \
  --set-env-vars="CLAUDE_API_KEY=${CLAUDE_API_KEY}"

Kubernetes Deployment

apiVersion: apps/v1
kind: Deployment
metadata:
  name: linkedin-agent
spec:
  replicas: 3
  selector:
    matchLabels:
      app: linkedin-agent
  template:
    metadata:
      labels:
        app: linkedin-agent
    spec:
      containers:
      - name: agent
        image: gcr.io/my-project/linkedin-agent:latest
        ports:
        - containerPort: 8080
        env:
        - name: CLAUDE_API_KEY
          valueFrom:
            secretKeyRef:
              name: agent-secrets
              key: claude-api-key
        resources:
          requests:
            memory: "512Mi"
            cpu: "500m"
          limits:
            memory: "1Gi"
            cpu: "1"

Best Practices

1. Agent Design

• Single Responsibility: Each agent should have one clear purpose
• Stateless: Agents should be stateless; use external storage for state
• Idempotent: Operations should be safe to retry
• Timeout: Always set timeouts for agent operations
• Rate Limiting: Respect API rate limits

2. Error Handling

class RobustAgent:
    async def run_with_retry(self, input_data):
        """Run with exponential backoff retry"""
        max_retries = 3
        base_delay = 1

        for attempt in range(max_retries):
            try:
                return await self.run(input_data)
            except RateLimitError as e:
                wait_time = base_delay * (2 ** attempt)
                logger.warning(f"Rate limited, waiting {wait_time}s")
                await asyncio.sleep(wait_time)
            except Exception as e:
                logger.error(f"Attempt {attempt + 1} failed: {e}")
                if attempt == max_retries - 1:
                    raise

3. Observability

from prometheus_client import Counter, Histogram
import structlog

# Metrics
agent_runs = Counter('agent_runs_total', 'Total agent runs')
agent_duration = Histogram('agent_duration_seconds', 'Agent run duration')

# Structured logging
logger = structlog.get_logger()

class ObservableAgent:
    @agent_duration.time()
    async def run(self, input_data):
        """Run with observability"""
        agent_runs.inc()

        logger.info(
            "agent_started",
            agent_name=self.__class__.__name__,
            input_size=len(str(input_data))
        )

        try:
            result = await self._execute(input_data)
            logger.info(
                "agent_completed",
                success=True,
                result_size=len(str(result))
            )
            return result
        except Exception as e:
            logger.error(
                "agent_failed",
                error=str(e),
                error_type=type(e).__name__
            )
            raise

4. Security

• API Key Management: Use Secret Manager, never hardcode
• Input Validation: Always validate and sanitize inputs
• Rate Limiting: Implement rate limiting for public endpoints
• Authentication: Use proper authentication for agent APIs
• Least Privilege: Grant minimal necessary permissions

Performance Tuning

Optimization Strategies

class OptimizedAgent:
    def __init__(self):
        # Connection pooling
        self.session = aiohttp.ClientSession(
            connector=aiohttp.TCPConnector(limit=100)
        )

        # Caching
        self.cache = TTLCache(maxsize=1000, ttl=3600)

        # Batch processing
        self.batch_size = 10
        self.queue = asyncio.Queue(maxsize=100)

    async def process_batch(self, items):
        """Process items in batches for efficiency"""
        tasks = []
        for batch in self.chunk(items, self.batch_size):
            tasks.append(self.process_chunk(batch))

        results = await asyncio.gather(*tasks)
        return self.flatten(results)

    @lru_cache(maxsize=128)
    def expensive_computation(self, input_data):
        """Cache expensive computations"""
        # Computation logic
        pass

Resource Management

# Resource limits for Kubernetes
resources:
  requests:
    memory: "256Mi"
    cpu: "250m"
  limits:
    memory: "1Gi"
    cpu: "1"

# Autoscaling
autoscaling:
  minReplicas: 1
  maxReplicas: 10
  targetCPU: 70
  targetMemory: 80

Integration Examples

With Docker Plugin

# Generate agent
adk-agent create --name my-agent --output ./my-agent/

# Containerize with Docker plugin
docker-agent create \
  --name my-agent \
  --base-image python:3.11-slim \
  --source ./my-agent/

With Terraform Plugin

# Generate infrastructure
terraform-gcp create \
  --name my-agent \
  --deploy-target cloud-run \
  --source ./my-agent/

With CI/CD

# GitHub Actions workflow
name: Deploy Agent
on:
  push:
    branches: [main]

jobs:
  deploy:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v3

      - name: Build Agent
        run: |
          adk-agent build --name my-agent

      - name: Run Tests
        run: |
          pytest tests/ --cov=agent --cov-report=xml

      - name: Build Container
        run: |
          docker build -t gcr.io/${{ secrets.GCP_PROJECT }}/my-agent:${{ github.sha }} .

      - name: Deploy to Cloud Run
        run: |
          gcloud run deploy my-agent \
            --image gcr.io/${{ secrets.GCP_PROJECT }}/my-agent:${{ github.sha }} \
            --region us-central1

Troubleshooting

Common Issues

Agent Not Responding

# Check logs
kubectl logs -f deployment/my-agent

# Check health
curl http://agent-service:8080/health

# Debug mode
AGENT_DEBUG=true python agent.py

High Latency

• Check Claude API response times
• Review tool execution times
• Enable caching for repeated operations
• Use batch processing where possible

Memory Issues

# Memory profiling
from memory_profiler import profile

@profile
def memory_intensive_function():
    # Your code
    pass

# Garbage collection tuning
import gc
gc.set_threshold(700, 10, 10)

Cost Optimization

Claude API Costs

class CostOptimizedAgent:
    def __init__(self):
        # Use smaller model for simple tasks
        self.models = {
            "simple": "claude-3-haiku-20240307",  # Cheapest
            "standard": "claude-3-5-sonnet-20241022",  # Balanced
            "complex": "claude-3-5-opus-20241022"  # Most capable
        }

    def select_model(self, task_complexity):
        """Select appropriate model based on task"""
        if task_complexity < 0.3:
            return self.models["simple"]
        elif task_complexity < 0.7:
            return self.models["standard"]
        else:
            return self.models["complex"]

Resource Optimization

• Use Cloud Run with scale-to-zero for development
• Implement request batching to reduce API calls
• Cache frequently accessed data
• Use preemptible VMs for batch processing

Examples Repository

Complete examples available at:

/examples/

• simple-react-agent/

  - Basic React pattern agent

• sdr-team/

  - Multi-agent SDR team

• enrichment-workflow/

  - Data enrichment pipeline

• real-time-agent/

  - WebSocket-based real-time agent

• batch-processor/

  - High-volume batch processing

Support

• GitHub Issues: https://github.com/jeremylongshore/jeremy-google-adk/issues
• Documentation: https://claudecodeplugins.io/plugins/jeremy-google-adk
• Community: Discord/Slack channel

---

Version: 1.0.0 Last Updated: October 2025 Author: Jeremy Longshore License: MIT