Google ADK (Agent Development Kit)

Analogy - Google Office Building

Imagine moving into a fully furnished Google office building. The WiFi, phones, meeting rooms, security system - everything is already set up and works together seamlessly. Need to search something? Google Search is built in. Need a map? Google Maps is right there. That is what ADK feels like - an agent framework where all Google services are already connected and ready to use.

ADK in Simple Terms:

Google ADK (Agent Development Kit) is Google's open-source framework for building AI agents. Released in April 2025, it is designed to work seamlessly with Gemini models and Google Cloud services. Key features:

• Multi-Agent Architecture: Build systems with multiple specialized agents that coordinate automatically
• Rich Tool Ecosystem: Built-in tools for Google Search, code execution, Maps, YouTube, and more
• Vertex AI Integration: Deploy agents to production on Google Cloud with enterprise features
• Evaluation Built-in: Test and evaluate agents without third-party tools
• Streaming & Voice: Real-time audio/video streaming for conversational agents

Not Just Gemini:

While ADK is optimized for Gemini, it supports other models too through LiteLLM integration. You can use GPT-4, Claude, or open-source models as the brain of your agent. However, you get the best experience with Gemini because of the deep integration with Google services.

Agent Types in ADK:

• LLM Agent: The standard agent powered by a language model. Has instructions, tools, and can reason about tasks. Like a smart employee who thinks and decides.
• Sequential Agent: Runs a series of sub-agents one after another in a fixed order. Like a factory assembly line - step 1, then step 2, then step 3.
• Parallel Agent: Runs multiple sub-agents at the same time. Like asking three different teams to research simultaneously and combining their results.
• Loop Agent: Repeats a sub-agent until a condition is met. Like a quality checker who keeps reviewing until the output is good enough.

Tool Categories:

• Built-in Tools: Google Search, Code Execution, Vertex AI Search - ready to use out of the box
• Function Tools: Your own Python functions wrapped as tools
• Third-Party Tools: Connect to external APIs through OpenAPI specs or MCP servers
• Agent Tools: Use another agent as a tool - powerful for hierarchical designs

Agent-as-Tool Pattern:

In ADK, you can use one agent as a tool for another. The parent agent decides when to call the child agent, passes the task, and gets results back. This is like a manager delegating to team members - the manager does not do the work but knows who to assign it to.

Session Management:

ADK has a built-in session system that manages conversations across multiple turns. Each session tracks:

• Conversation history (all messages)
• Session state (key-value data that persists across turns)
• Active agent (which agent is currently handling the conversation)

You can store sessions in memory (development), a database (production), or Google Cloud storage (enterprise).

State Management:

ADK provides three scopes of state:

• App State: Shared across all users and sessions. Good for global config.
• User State: Per-user data that persists across sessions. Good for preferences.
• Session State: Per-session data. Good for current conversation context.

Memory Service:

Beyond session state, ADK offers a Memory Service for long-term memory. It can remember facts from past conversations and retrieve them when relevant. Like how a good doctor remembers your medical history across visits. This uses Vertex AI Search or custom memory backends.

Travel Planning Agent System:

Root Agent (LLM Agent - Gemini 2.0 Flash)
  "You are a travel planning assistant for Indian travelers"
  |
  +-- Research Agent (Parallel Agent)
  |   Runs these three simultaneously:
  |   |-- Flight Agent: Search flights via Skyscanner API
  |   |-- Hotel Agent: Search hotels via booking API  
  |   |-- Activity Agent: Find activities via Google Search
  |
  +-- Budget Agent (LLM Agent)
  |   "Calculate total trip cost in INR, suggest money-saving tips"
  |   Tools: [currency_converter, budget_calculator]
  |
  +-- Booking Agent (Sequential Agent)
      Step 1: Verify availability
      Step 2: Get user confirmation
      Step 3: Make bookings
      Step 4: Send confirmation email

How It Works:

1. User: "Plan a 3-day Goa trip for 2 people, budget Rs 30,000"
2. Root Agent understands the request, delegates to Research Agent
3. Research Agent runs Flight, Hotel, and Activity agents in parallel (saves time)
4. Results come back, Root Agent sends to Budget Agent for cost analysis
5. Budget Agent suggests best combinations within Rs 30,000 budget
6. User approves, Booking Agent handles the sequential booking flow

ADK Advantage:

Notice how ADK's agent types (Parallel, Sequential, LLM) naturally map to real workflows. Research happens in parallel (faster), booking happens sequentially (order matters). You express this directly in code without writing complex orchestration logic.

Google Ecosystem Bias:

ADK works best within the Google ecosystem. While it supports other models via LiteLLM, the tooling, deployment, and integrations are optimized for Gemini + Google Cloud + Vertex AI. If you are on AWS or Azure, the deployment story is less smooth.

Newer Framework:

ADK was released in April 2025, making it relatively new compared to LangChain or even OpenAI Agents SDK. The community is growing but smaller. Documentation and tutorials are still maturing. Expect more breaking changes as the API evolves.

When to Choose ADK:

• Choose ADK if: You use Google Cloud, need Gemini integration, want built-in evaluation, need structured multi-agent workflows
• Choose OpenAI SDK if: You are committed to OpenAI models and want the simplest possible setup
• Choose Claude/MCP if: You want model flexibility, extended thinking, and the MCP ecosystem
• Choose LangChain if: You need maximum flexibility and the largest community

Q: What are the four agent types in Google ADK and when do you use each?

Q: How does ADK handle state management?

Q: How does Google ADK compare to OpenAI Agents SDK?