Google ADK
Google · Google Agent Development Kit (ADK): Open-Source Multi-Language Agent Framework
Google's Agent Development Kit (ADK) is an open-source framework for building, evaluating, and deploying AI agents and multi-agent systems, with SDKs for Python, TypeScript, Go, Java, and Kotlin. It is aimed at developers who want code-first agents that are optimized for Gemini but can run other models and deploy to Google Cloud or their own infrastructure.
Best for
Developers building production agents in Python, TypeScript, Go, Java, or Kotlin who want an open-source, code-first framework with first-class Gemini support and a direct path to Google Cloud
Not ideal for
Non-developers looking for a no-code agent builder, or teams that want identical features across every language SDK today
Who it's for
Software engineers building production AI agents and multi-agent systems, especially teams using Gemini or Google Cloud
Capabilities
- Official SDKs for Python, TypeScript, Go, Java, and Kotlin, all under Apache-2.0
- Graph workflows (ADK 2.0, in Python, TypeScript, and Go) that combine deterministic code paths with model-driven reasoning
- Multi-agent composition with sequential, parallel, and loop patterns, agent routing, and hierarchical agents
- Built-in tools for Google Search, OpenAPI specs, and MCP, plus custom function tools with action confirmations
- Context management that filters, summarizes, and lazy-loads context and tracks token usage
- `adk web` developer UI for running and debugging agents, with an experimental Visual Builder (Python) that generates agent configs and code
- Evaluation tools (Python) with criteria, user simulation, environment simulation, and custom metrics
- Experimental Agent2Agent (A2A) protocol support for exposing ADK agents and consuming remote agents
- Model support for Gemini and Gemma plus Claude, OpenAI, Ollama, vLLM, and LiteLLM, and on-device Android agents via LiteRT-LM and ML Kit (beta) in ADK for Kotlin
- Deploy anywhere as a container, or to Google Cloud Agent Runtime, Cloud Run, or GKE
Limitations
- Feature coverage differs by language: graph workflows cover Python, TypeScript, and Go; evaluation and the Visual Builder are Python-only; A2A docs cover Python, Go, and Java; and the Agent Runtime and GKE deployment guides cover Python and Go
- The Visual Builder and A2A support are labeled experimental
- Optimized for Gemini; other providers are supported through model integrations
- The managed deployment path runs through Google Cloud, and Agent Runtime is a paid service above its no-cost tier
- Code-first framework that requires programming; it is not a hosted no-code agent builder
Use cases
- Building a multi-agent customer workflow in Python where a router agent delegates to specialist sub-agents
- Adding an agent to an existing Go, Java, or TypeScript backend using the native SDK instead of a Python sidecar
- Shipping an Android app with an on-device agent that uses LiteRT-LM for private, local inference
- Exposing an internal agent over A2A so agents built by other teams can call it
- Evaluating agent behavior with user simulation and custom metrics before deploying to Cloud Run or Agent Runtime
Our take
ADK is Google's first-party framework, so it is the most direct option for teams building agents on Gemini or Google Cloud. Its language coverage also matters outside Google: official Go, Java, and Kotlin SDKs sit alongside Python and TypeScript, and ADK for Kotlin adds on-device Android agents. Graph workflows, evaluation, and A2A support cover much of what production teams need. The caveats are uneven feature coverage across languages (evaluation and the Visual Builder are Python-only, and A2A is experimental) and a managed deployment story centered on Google Cloud. Python-only teams not on Google Cloud may also want to compare LangGraph or Pydantic AI.
Who should use it
Teams building on Gemini or Google Cloud, backend engineers who want agents in Go, Java, TypeScript, or Kotlin without a Python service, Android developers exploring on-device agents, and teams that need A2A interoperability between agents.
Who should skip it
Non-developers who want a no-code builder, teams that need every feature in a language other than Python today, and organizations that want a fully managed hosted agent platform outside Google Cloud.
Strengths
- Official SDKs in five languages, including Go, Java, and Kotlin
- Open source under Apache-2.0 with no framework fees
- Graph workflows, evaluation tooling, and A2A support (experimental) in one framework
- Works with Gemini plus Claude, OpenAI, and local models through Ollama or vLLM
- On-device Android agents through ADK for Kotlin
Weaknesses
- Features vary across language SDKs; evaluation and the Visual Builder are Python-only
- A2A support and the Visual Builder are experimental
- Managed deployment is built around Google Cloud
- Requires coding, so it is not suited to no-code users
Google ADK pricing
Open Source
Free
- Apache-2.0 SDKs for Python, TypeScript, Go, Java, and Kotlin
- Run on your own infrastructure
Note: The framework itself is free. Costs come from the model providers you call and from hosting. Google Cloud Agent Runtime is a paid service beyond its no-cost access tier.
Technical specs
Available models
Where Google ADK excels
Adding agents to a Go or Java microservice stack
The official Go and Java SDKs let teams build agents in the same language as their existing services instead of running a separate Python service.
Building a privacy-sensitive assistant inside an Android app
ADK for Kotlin runs on-device agents through LiteRT-LM and ML Kit (beta), keeping inference on the phone when needed and using Firebase AI for cloud reasoning otherwise.
Connecting agents owned by different teams
ADK's experimental A2A support lets one team expose an ADK agent as a server and other agents call it as a remote agent, passing reasoning traces and file artifacts between them.
Google ADK vs. competitors
Google ADK vs. LangGraph
LangGraph models agent workflows as explicit graphs in Python and JS/TS. ADK also offers graph workflows in 2.0 (Python, TypeScript, Go) and adds official Go, Java, and Kotlin SDKs, built-in evaluation in Python, and a direct deployment path to Google Cloud.
Google ADK vs. Microsoft Agent Framework
Microsoft Agent Framework covers Python and .NET (Go in preview) and fits naturally with Azure. ADK covers Python, TypeScript, Go, Java, and Kotlin and fits naturally with Gemini and Google Cloud. The choice often follows your cloud and language stack.
Google ADK vs. CrewAI
CrewAI focuses on role-based teams of agents in Python with an optional hosted platform. ADK is a lower-level, multi-language framework with graph workflows, A2A support, and evaluation tooling.
Google ADK vs. Pydantic AI
Pydantic AI is a Python framework centered on type-safe, validated structured outputs. ADK is broader in scope (multi-agent orchestration, deployment, evaluation) and spans five languages.
Frequently asked questions
What is Google ADK?
The Agent Development Kit (ADK) is Google's open-source framework for building, debugging, evaluating, and deploying AI agents and multi-agent systems. It has official SDKs for Python, TypeScript, Go, Java, and Kotlin.
Is Google ADK free and open source?
Yes. The ADK SDKs are open source under the Apache-2.0 license. You pay for the models you call and for any hosting, such as Google Cloud Agent Runtime, which is a paid service above its no-cost tier.
Does ADK only work with Gemini?
No. ADK is optimized for Gemini but supports Gemma, Anthropic Claude, OpenAI models, Ollama, vLLM, and LiteLLM, and the Kotlin SDK can run on-device models through LiteRT-LM.
Where can I deploy ADK agents?
ADK agents can be containerized and run on your own infrastructure, or deployed to Google Cloud through Agent Runtime (formerly Vertex AI Agent Engine), Cloud Run, or GKE.
Can I build ADK agents for Android?
Yes. ADK for Kotlin 1.0, released in September 2026, includes Android extensions for on-device agents using LiteRT-LM and ML Kit (beta), plus Firebase AI integration for cloud-backed reasoning.
Integrations & fit
Alternatives to consider
LangChain
Developers building production multi-agent systems that need fine-grained control over state, execution flow, and human-in-the-loop checkpoints — and who are willing to trade setup time for that control
FreeMicrosoft
Python and .NET multi-agent orchestration, especially with Azure and Microsoft Foundry
FreeCrewAI
Orchestrating autonomous agent teams for enterprise tasks
FreemiumPydantic
Type-safe Python agents with validated structured outputs
FreeAbout Google ADK
ADK treats agent development like ordinary software engineering. You define agents in code, compose them into hierarchies or sequential, parallel, and loop workflows, and, in ADK 2.0, mix deterministic graph workflows with model-driven reasoning. It comes with built-in tools (Google Search, OpenAPI, MCP, and custom function tools with confirmation steps), context management that filters, summarizes, and lazy-loads context, and an evaluation suite with criteria, user simulation, and custom metrics. The `adk web` developer UI runs and debugs agents locally. A Visual Builder inside it generates agent configs and Python code, but it is experimental and Python-only. For multi-agent systems across services, ADK has experimental support for the Agent2Agent (A2A) protocol to expose agents and call remote ones. ADK ships official SDKs for Python, TypeScript (2.0 GA), Go, Java, and Kotlin, but features differ by language: Python has the broadest coverage, graph workflows cover Python, TypeScript, and Go, and the evaluation tooling and Visual Builder are Python-only. ADK for Kotlin 1.0 (September 2026) adds on-device Android agents through LiteRT-LM and ML Kit (beta). ADK is optimized for Gemini but also supports Gemma, Claude, OpenAI models, Ollama, vLLM, and LiteLLM. Deployment is flexible: you can containerize agents anywhere or deploy to Google Cloud's Agent Runtime (formerly Vertex AI Agent Engine), Cloud Run, or GKE. The framework is free under Apache-2.0. Costs come from model usage and from Agent Runtime, which is paid above its no-cost tier. Tradeoffs: features are not identical across language SDKs, the smoothest managed path runs through Google Cloud, and it is a code-first framework rather than a no-code builder.
Updates from Google ADK
Google released ADK for Kotlin 1.0 with feature parity with ADK 1.0 Core, a Kotlin Multiplatform core, and Android extensions for on-device agents through LiteRT-LM and ML Kit (beta), plus Firebase AI integration and persistence via Room and AppSearch.
Are you the founder? Claim this listing →