adk-engineer

name	adk-engineer
description	Implement production-ready agents using Google Agent Development Kit (ADK). Use when creating ADK agents, building multi-agent systems, implementing workflow agents, reviewing ADK code, or deploying to Vertex AI / Cloud Run. Triggers: 'ADK agent', 'Google ADK', 'multi-agent system', 'Gemini agent', 'agent workflow', 'adk deploy'. Do NOT use for non-ADK agent frameworks (LangChain, CrewAI, AutoGen) or general Gemini API calls without ADK.
allowed-tools	["Read","Edit","Write","Bash","Grep","Glob","WebFetch"]

ADK Engineer

Implement, review, and deploy Google ADK agents following proven patterns and best practices.

Quick Start

Single agent: Read references/adk-patterns.md -> Use assets/templates/single-agent/ -> Validate with references/adk-best-practices.md.

Multi-agent system: Choose pattern (Coordinator, Sequential, Parallel, Loop) -> Use matching template from assets/templates/ -> Consult references/adk-patterns.md.

Pattern Selection

Task Type	Pattern	Template
Linear data processing	Sequential Pipeline	assets/templates/workflow/
Intent-based routing	Coordinator/Dispatcher	assets/templates/multi-agent/
Independent concurrent tasks	Parallel Fan-Out/Gather	assets/templates/workflow/
Nested agent structures	Hierarchical Decomposition	assets/templates/multi-agent/
Validation loops	Generator-Critic	assets/templates/workflow/
Quality improvement	Iterative Refinement	assets/templates/workflow/
Critical decisions	Human-in-the-Loop	assets/templates/single-agent/

Implementation Workflow

New Agent

1. Choose pattern from table above based on task analysis.
2. Copy template:

   cp -r assets/templates/[single-agent|multi-agent|workflow]/ ./my_agent/
   # Or: bash scripts/init_adk_project.sh my_agent [template_type]
   

3. Implement domain logic: Define custom tools, update instructions, configure output_key values.
4. Test locally: Run python agent.py or adk ui.
5. Validate:

   python scripts/validate_agent.py agent.py
   

6. Deploy per references/adk-deployment.md.

Code Review

1. Read agent code and identify patterns used.
2. Check implementation against references/adk-patterns.md.
3. Review against references/adk-best-practices.md.
4. Run python scripts/validate_agent.py agent.py.
5. Provide specific, actionable recommendations with references.

Validation Checklist

• root_agent properly defined
• Agent descriptions are clear and specific
• Descriptive output_key values for state management
• Loop agents have max_iterations
• Parallel agents use unique output_key values
• Tools follow verb-noun naming pattern
• Error handling with graceful degradation
• Appropriate model selection per agent complexity
• REQUIRE_CONFIRMATION for high-stakes operations

Deployment

# Vertex AI Agent Engine (recommended)
adk deploy agent_engine --project=PROJECT_ID --region=us-central1 \
  --staging_bucket=gs://BUCKET/staging --agent-file=agent.py

# Cloud Run
adk deploy cloud_run --project=PROJECT_ID --region=us-central1 \
  --agent-file=agent.py

See references/adk-deployment.md for Docker, CI/CD, monitoring, and security details.

Error Handling

• If validate_agent.py reports issues, fix all CRITICAL/HIGH items before deployment.
• If pattern selection is ambiguous, default to Coordinator for routing tasks, Sequential for pipelines.
• If deployment fails, check project permissions and staging bucket access first.

References

• references/adk-patterns.md -- 7 design patterns with code examples
• references/adk-best-practices.md -- Agent design, state management, tools, security
• references/adk-deployment.md -- Deployment workflows, CI/CD, monitoring, cost optimization
• assets/templates/ -- Single-agent, multi-agent, workflow starter templates
• Google ADK Docs