Google Agent Development Kit (ADK): Complete Platform Profile

Google Agent Development Kit (ADK) is Google's open-source framework for building production-grade AI agents and multi-agent systems. Announced at Google Cloud Next 2025, ADK reflects Google's strategy of providing a complete developer toolkit that spans from local prototyping through large-scale cloud deployment. While optimized for Gemini models and Google Cloud infrastructure, the framework is explicitly designed to be model-agnostic and cloud-agnostic, supporting deployment on any cloud or on-premises environment.

Visit the AI agent tools directory to compare Google ADK against the full spectrum of agent frameworks available to developers today.

Overview#

Google released ADK in April 2025 as part of a broader push to establish leadership in the enterprise agent market. The timing coincided with significant improvements to the Gemini model family, particularly around long context windows and function calling reliability, which are prerequisites for effective agent operation.

ADK draws on lessons from Google's internal experience deploying agents at scale. The framework was used internally before being released publicly, which gives it a degree of production-hardening that purely community-developed frameworks sometimes lack. Google has been transparent that ADK powers several of its own AI products, including Google Workspace AI assistants.

A key technical innovation in ADK is its support for the Agent2Agent (A2A) protocol — an open communication standard that allows agents built with different frameworks to interoperate. An agent built with ADK can receive tasks from a CrewAI orchestrator, delegate subtasks to an AutoGen agent, and return results to a LangGraph workflow, all through a standardized API. This interoperability focus positions ADK within a larger industry conversation about avoiding agentic lock-in.

The framework has garnered significant attention from enterprises already invested in Google Cloud, for whom ADK represents a natural extension of their existing infrastructure and tooling relationships.

Core Features#

Hierarchical Multi-Agent Orchestration#

ADK is built around the concept of agent hierarchies. A root orchestrator agent decomposes complex tasks and delegates subtasks to specialized sub-agents. Those sub-agents can themselves delegate to lower-level agents, creating trees of cooperation that can tackle problems of arbitrary complexity.

This hierarchical model is implemented through ADK's routing logic, which supports several delegation patterns: LLM-driven routing (the orchestrator uses a language model to decide which sub-agent to call), rule-based routing (conditions determine routing without a model call), and explicit parallel execution (multiple sub-agents run concurrently and their results are merged). Developers can mix these patterns within the same hierarchy depending on the latency and reliability requirements of each node.

The orchestration layer handles context propagation automatically. When an orchestrator delegates to a sub-agent, ADK manages how much conversation history and shared state to include, configurable via context window management policies. This prevents sub-agents from receiving irrelevant context that wastes tokens and degrades performance.

Agent2Agent (A2A) Protocol#

ADK's support for the A2A protocol is one of its most strategically significant features. A2A is an open standard developed collaboratively by Google and other industry participants that defines how agents communicate — how they negotiate capabilities, exchange tasks, report progress, and return results. It operates over standard HTTP and uses a JSON-based message format, making it implementable in any language.

In practice, A2A support means ADK agents can operate as either A2A clients (delegating tasks to other agents) or A2A servers (exposing themselves as callable services). An organization can build a distributed agentic system where specialized agents — perhaps developed by different teams using different frameworks — coordinate through the A2A protocol without requiring shared infrastructure or a common codebase.

This interoperability story is particularly compelling for large enterprises with diverse technology stacks and AI initiatives across multiple business units.

Built-in Tools and MCP Support#

ADK ships with a substantial library of built-in tools covering common agent needs: Google Search integration, code execution via a sandboxed Python interpreter, file system access, and connectors to Google Workspace APIs. These tools are implemented to the same interface as user-defined tools, meaning they can be composed freely with custom functionality.

ADK also supports the Model Context Protocol (MCP), the standard for tool discovery and execution popularized by Anthropic. An ADK agent can connect to any MCP server and use its tools without custom integration code. This dramatically expands the effective tool library available to ADK agents, since the MCP ecosystem includes hundreds of community-built servers for databases, SaaS platforms, developer tools, and data sources.

Vertex AI Integration and Deployment#

For teams deploying on Google Cloud, ADK's Vertex AI integration provides managed infrastructure for agent execution. Agents deployed to Vertex AI Agent Builder receive automatic scaling, built-in logging, IAM-based access control, and integration with Google Cloud's observability stack (Cloud Logging, Cloud Monitoring, Cloud Trace).

The integration also enables access to Vertex AI's specialized capabilities: grounding with Google Search, document understanding APIs, and vector search through Vertex AI Feature Store. These capabilities extend what agents can do beyond what's possible with raw model calls.

Pricing and Plans#

ADK itself is free and open source under the Apache 2.0 license. Running ADK agents incurs costs only through the services they consume. For agents using Gemini models, Google Cloud charges per token through the Vertex AI API or the AI Studio API. Gemini 2.0 Flash offers a particularly cost-effective option for high-volume agent workloads.

Managed deployment through Vertex AI Agent Builder introduces additional platform costs, priced based on conversation volume and compute hours. Google Cloud's pricing calculator provides estimates for specific usage patterns. Enterprise customers can negotiate committed-use discounts.

There is no charge for the A2A protocol implementation or MCP integration capabilities.

Strengths#

First-class Google Cloud integration. For organizations already running on Google Cloud, ADK reduces integration friction for agent deployment dramatically. IAM, logging, monitoring, and scaling are handled by familiar managed services.

A2A protocol support enables genuine interoperability. The ability to compose agents across frameworks through a standardized protocol is a meaningful technical differentiator that no other major framework currently offers at the same level.

Built-in code execution sandbox. The ability to run Python code safely in a sandboxed environment, built directly into the framework's tool library, is valuable for data analysis agents, automation agents, and any workflow that benefits from programmatic computation.

Strong support for Gemini's unique capabilities. Gemini's long context window (up to 1 million tokens), native multimodal inputs, and advanced function calling are all surfaced through ADK's abstractions in ways that take full advantage of the model family's differentiated capabilities.

Limitations#

Google Cloud affinity creates implicit lock-in. While ADK is technically provider-agnostic, many of its most compelling features — Vertex AI deployment, Workspace integration, grounding — only work within the Google Cloud ecosystem. Teams that need true cloud portability may find this friction significant.

Relatively young ecosystem. ADK does not yet have the depth of community-contributed integrations, tutorials, and production case studies that older frameworks like LangChain possess. Documentation, while good, has gaps in edge cases.

Complexity of multi-agent setup. The power of ADK's hierarchical orchestration comes with setup complexity. Getting a multi-agent hierarchy right requires understanding routing policies, context propagation, and the A2A handshake process in detail.

Ideal Use Cases#

• Enterprise automation on Google Cloud: Deploy multi-agent workflows that leverage Google Workspace APIs, BigQuery, and Vertex AI as components in a unified orchestration layer.
• Cross-framework multi-agent systems: Build distributed agentic architectures where ADK agents coordinate with agents built in other frameworks through the A2A protocol.
• Research and analysis agents: Use ADK's code execution sandbox and long-context Gemini models to build agents that write, execute, and iterate on data analysis code.
• Document intelligence pipelines: Process large document collections using Gemini's extended context and Vertex AI's document understanding APIs through ADK's orchestration layer.

Getting Started#

Install ADK and authenticate with Google Cloud:

pip install google-adk
gcloud auth application-default login

Create a simple agent with a built-in tool:

from google.adk.agents import Agent
from google.adk.tools import google_search

root_agent = Agent(
    name="research_agent",
    model="gemini-2.0-flash",
    description="An agent that can research topics using Google Search.",
    instruction="You are a helpful research assistant. Use Google Search to find accurate, current information.",
    tools=[google_search],
)

To run the agent interactively using ADK's web UI for local development:

adk web

For multi-agent hierarchies, define sub-agents first, then reference them in the orchestrator's tools list. The ADK documentation provides complete examples for LLM-driven routing, parallel delegation, and A2A server deployment.

How It Compares#

Google ADK vs LangChain/LangGraph: LangGraph's graph-based orchestration gives more fine-grained control over agent state machines. ADK's hierarchical model is higher-level and faster to configure for standard delegation patterns. LangGraph suits teams that need precise control over complex conditional flows; ADK suits teams that need speed and Google Cloud integration.

Google ADK vs OpenAI Agents SDK: The OpenAI Agents SDK is tighter and simpler, optimized for OpenAI's specific capabilities. ADK is broader in scope and explicitly multi-cloud. The choice largely follows your primary model vendor relationship.

Google ADK vs Semantic Kernel: Semantic Kernel is Microsoft's enterprise framework with strong Azure integration — the direct counterpart in the Microsoft ecosystem. If your infrastructure is primarily Azure, Semantic Kernel is the more natural fit; if it's primarily Google Cloud, ADK is.

Bottom Line#

Google ADK is a serious, production-oriented framework that reflects genuine investment in the agent infrastructure problem. Its A2A protocol support is genuinely differentiated, its Google Cloud integration is deep, and its built-in tools make many common agent patterns easy to implement without custom code.

The framework will resonate most strongly with organizations already invested in Google Cloud and Gemini. For those organizations, ADK removes most of the integration work that would otherwise be required to deploy agents into a managed, observable, and scalable production environment.

Best for: Engineering teams deploying multi-agent systems on Google Cloud who want first-party framework support, Gemini model optimization, and the ability to compose across agent frameworks through the A2A protocol.

Frequently Asked Questions#

What is the Agent2Agent (A2A) protocol? A2A is an open communication standard that allows agents built with different frameworks to interoperate. It defines how agents negotiate capabilities, delegate tasks, and return results using standard HTTP and JSON. ADK agents can act as both A2A clients and servers. The protocol is designed to prevent agentic lock-in by making agents framework-agnostic at the communication level.

Does Google ADK work with models other than Gemini? Yes. ADK is designed to be model-agnostic and supports OpenAI-compatible APIs and Anthropic's Claude. However, the deepest integrations and most complete tooling are oriented around Gemini and Vertex AI.

Can I use Google ADK without a Google Cloud account? Yes. ADK can be used with the Google AI Studio API, which has a free tier, or with other model providers. The managed deployment features require Google Cloud, but local development and testing do not.

How does ADK handle long-running tasks? ADK supports asynchronous task execution and session management, allowing agents to pause and resume work across multiple interactions. For very long-running operations, the Vertex AI deployment infrastructure provides durability guarantees that local execution cannot.

What is the relationship between ADK and Google's Dialogflow? ADK and Dialogflow serve different needs. Dialogflow is a managed conversational AI platform with visual tooling aimed at business users. ADK is a developer framework for building custom agents with full programmatic control. They can be used together, but they are not alternatives to each other.