Google Agent Development Kit
Last reviewed
Sources
20 citations
Review status
Source-backed
Revision
v1 ยท 2,034 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
20 citations
Review status
Source-backed
Revision
v1 ยท 2,034 words
Add missing citations, update stale details, or suggest a clearer explanation.
The Google Agent Development Kit (ADK) is an open-source framework from Google for building, evaluating, and deploying AI agents and multi-agent systems. Google announced it on April 9, 2025 at Google Cloud Next 2025 and released it under the Apache 2.0 license, first as a Python library, with Java, Go, and TypeScript implementations following over the next year. ADK is model-agnostic but is tuned for Google's Gemini models and for deployment to the Google Cloud and Vertex AI stack. It is the same framework Google uses internally to build the agents inside products such as Agentspace (later rebranded Gemini Enterprise) and its Customer Engagement Suite. ADK ships alongside a companion open protocol, the Agent2Agent (A2A) protocol, for agent-to-agent interoperability, and it supports Anthropic's Model Context Protocol for connecting agents to tools and data.
| Attribute | Detail |
|---|---|
| Developer | Google / Google Cloud |
| Initial release | April 9, 2025 (Google Cloud Next 2025) |
| License | Apache 2.0 (open source) |
| Languages | Python, Java (and Kotlin on the JVM), Go, TypeScript/JavaScript |
| Python milestones | v1.0.0 stable May 20, 2025; v2.0.0 GA May 19, 2026; v2.2.0 June 4, 2026 |
| Default model | Gemini; model-agnostic via LiteLLM |
| Companion protocol | Agent2Agent (A2A); also supports MCP |
| Managed runtime | Vertex AI Agent Engine (also Cloud Run, GKE, or self-hosted) |
| Repositories | google/adk-python, google/adk-java, google/adk-go, google/adk-js |
| Samples hub | Agent Garden / google/adk-samples |
ADK is a "code-first" toolkit: developers define an agent's instructions, tools, and orchestration in ordinary source code rather than through a visual builder or a long prompt. Google's stated goal was to make agent development feel more like conventional software engineering, with version control, unit testing, and continuous integration, so that teams can move from a single agent with a few tool calls to large multi-agent systems without changing frameworks. The kit is deliberately positioned for production rather than demos; the official documentation site carries the tagline "Build production agents, not prototypes."
Although ADK works with many large language models, it is optimized for Gemini and for Google Cloud. Out of the box it gives direct access to Gemini through Vertex AI or the Gemini API, native generative AI tooling such as Google Search grounding and code execution, and one-command deployment to Google's managed agent runtime. The same agent code can also run locally, in a container, or on a competing cloud, which Google emphasizes to avoid the impression of lock-in.
ADK centers on composable agents. The core abstraction is an Agent (or LlmAgent) defined by a model, an instruction, a set of tools, and optionally a list of sub-agents. Key capabilities include:
AgentEvaluator.evaluate() in Python) that scores an agent's final responses and its step-by-step trajectory against predefined test cases, runnable from the CLI, a web UI, or programmatically.ADK began as Python (the google-adk package, installed with pip install google-adk). Google released the Python ADK v1.0.0 as a stable, production-ready version on May 20, 2025 at Google I/O, and shipped a 2.0 generation at Google I/O in May 2026 (v2.0.0 reached general availability on May 19, 2026, with point releases such as v2.2.0 on June 4, 2026). The Python repository has gathered more than 20,000 stars on GitHub.
The framework then expanded to other languages with the explicit aim of feature parity across SDKs:
| ADK implementation | First release | Notes |
|---|---|---|
| Python (adk-python) | April 9, 2025 | Reference implementation; v1.0.0 May 20, 2025 |
| Java (adk-java) | v0.1.0, May 20, 2025 | Announced at Google I/O 2025; v1.0.0 March 30, 2026 |
| Go (adk-go) | November 2025 | Adds Go to the JVM and Python SDKs |
| TypeScript/JavaScript (adk-js) | v1.0.0, April 21, 2026 | Node.js and browser runtimes, Zod schema validation |
The official docs site also lists Kotlin, which runs on the JVM alongside the Java SDK. This multi-language strategy lets backend, mobile, and web teams build agents in the language they already use.
Two interoperability layers sit around ADK. The first is the Agent2Agent (A2A) protocol, an open standard Google announced on the same day as ADK, April 9, 2025, with more than 50 launch partners. A2A lets agents built by different vendors or in different frameworks discover one another and collaborate. It is built on familiar web standards (HTTP, Server-Sent Events, and JSON-RPC) and uses "Agent Cards," JSON capability descriptions that let one agent learn what another can do before delegating a task. Google frames A2A as complementary to MCP: MCP connects an agent to tools and data, while A2A connects agents to other agents. In June 2025 Google donated A2A to the Linux Foundation, which launched the Agent2Agent project on June 23, 2025 with founding members including Amazon Web Services, Cisco, Google, Microsoft, Salesforce, SAP, and ServiceNow; backing grew to more than 100 companies, and over 150 organizations by 2026. ADK gained native A2A support so that ADK agents can act as A2A clients and servers.
The second layer is deployment. Google's managed runtime, Vertex AI Agent Engine, runs ADK agents without the developer managing servers or containers, with built-in session management, tracing, and enterprise security. The ADK CLI deploys to it with a single adk deploy command, which packages the agent, wraps it in an AdkApp, and creates the deployment through the agent_engines.create() API. Agent Engine added support for ADK v1.0.0 and later in the google-cloud-aiplatform package version 1.95.0 and up. Agents can equally be containerized and run on Cloud Run, Google Kubernetes Engine, or any other infrastructure. Rounding out the ecosystem is Agent Garden, a curated collection of ready-to-use sample agents and tools accessible from the docs and the google/adk-samples repository, with one-click deployment of samples to Agent Engine via the Agent Starter Pack.
ADK is woven through Google's own enterprise agent stack. Google has said publicly that ADK is the framework powering agents inside Agentspace (rebranded Gemini Enterprise) and the Google Customer Engagement Suite, so the kit's internal use is substantial. Google has cited early external users including Renault Group, the content platform Box, and the retail-pricing firm Revionics, and A2A launch and partner agents have come from companies such as Box, Workday, Salesforce, and ServiceNow. Independent developer coverage from outlets including InfoQ, InfoWorld, and DevOps.com tracked each language release through 2025 and 2026, and the Python repository's 20,000-plus GitHub stars indicate broad community uptake. Because much enterprise usage is private, precise deployment counts are not public, and figures here are limited to what Google and reporters have stated.
ADK entered a crowded field of agent frameworks that emerged in 2024 and 2025, and comparisons usually turn on orchestration style and ecosystem lock-in rather than raw capability:
| Framework | Maker | Orchestration model | Model support |
|---|---|---|---|
| Google ADK | Hierarchical agent tree plus workflow/graph agents | Gemini-first, model-agnostic via LiteLLM | |
| LangGraph | LangChain | Explicit directed graph with checkpointing | Model-agnostic |
| CrewAI | CrewAI | Role-based "crews" of agents | Model-agnostic |
| OpenAI Agents SDK | OpenAI | Agent handoffs | OpenAI-centric |
| AutoGen | Microsoft | Conversational group chat among agents | Model-agnostic |
OpenAI released its Agents SDK in March 2025, about a month before ADK. Relative to LangGraph, which is favored for fine-grained control and observability, and CrewAI, which is praised for a gentle learning curve in prototyping, ADK's distinguishing pitch is tight integration with Google Cloud: easy access to Gemini, native enterprise connectors, built-in evaluation, and a one-command path to a managed runtime. Unlike OpenAI's SDK, ADK is not tied to one model provider; through the LiteLLM integration it can drive models from Anthropic, Meta, Mistral, and others. The frameworks are not strictly rivals at the protocol layer, since A2A and MCP are designed to let agents built in any of them interoperate.
ADK is the developer-facing centerpiece of Google's broader enterprise agentic AI strategy. Coupled with the A2A protocol for interoperability, Vertex AI Agent Engine for managed hosting, Agent Garden for samples, and Gemini Enterprise as the end-user surface, it gives Google a full-stack answer to competing agent platforms from OpenAI, Anthropic, and Microsoft. Two choices stand out. First, Google open-sourced both the kit (Apache 2.0) and the A2A protocol, then handed A2A to the Linux Foundation, a standards play intended to make cross-vendor agent collaboration vendor-neutral rather than proprietary. Second, by using the same framework internally that it ships to customers, Google positioned ADK as battle-tested production infrastructure rather than a demo toolkit. The rapid expansion from one language to four within roughly a year, and the move to a 2.0 generation with graph-based workflows, reflect how quickly the agent-framework category matured between 2025 and 2026, and how central these kits have become to enterprise adoption of generative AI.