Google Agent Development Kit

What is Google Agent Development Kit?

Google Agent Development Kit (ADK) is Google's open-source, code-first framework for building, evaluating and deploying AI agents and multi-agent systems, first announced at Google Cloud NEXT in April 2025 and now shipped as four separately-versioned, production-grade SDKs — Python, Go, Java and TypeScript — plus an experimental Kotlin/Android runtime. It is optimized for, but not limited to, Google's Gemini models: through a LiteLLM integration it can also run OpenAI, Anthropic Claude, Cohere, local Ollama and 100+ other providers. Released under the Apache 2.0 licence (google/adk-python: 20,900+ stars), it ships native support for both dominant 2026 agent-interoperability standards — the Agent2Agent (A2A) protocol, which Google open-sourced under Linux Foundation governance, and the Model Context Protocol (MCP) — plus a free, self-hostable core and an optional managed hosting layer (Agent Runtime, part of the Gemini Enterprise Agent Platform) for teams that don't want to operate the infrastructure themselves.

What does Google Agent Development Kit do?

You define an agent's instructions, tools and model in code, compose multiple agents into a workflow using ADK 2.0's graph-based Workflow Runtime (sequential, parallel, loop and conditional routing, with retries and human-in-the-loop steps) or its Task API for structured multi-turn delegation between agents, and debug the whole thing locally with the built-in ADK Web UI (adk web) before shipping. An evaluation framework (adk eval) lets you write repeatable test cases that score an agent's trajectory and final response, not just spot-check outputs by hand. For production, you can containerize and self-host the agent anywhere — Cloud Run, GKE, Docker, or any other infrastructure — for free, paying only your model provider; or deploy to the managed Agent Runtime (formerly branded Vertex AI Agent Engine) for autoscaling, session/memory persistence and built-in observability, billed on Google Cloud's usual per-vCPU/per-GB-hour metering. Agents can also connect to a low-code visual builder (Agent Designer, part of the wider Gemini Enterprise Agent Platform) for teams that want to configure simpler agents without writing code alongside ADK's own code-first agents.

How Google Agent Development Kit works

1. 1Define an agent's instructions, tools (plain functions or MCP-connected tools) and model — native Gemini, or any other provider via LiteLLM.
2. 2Compose multiple agents into a workflow using the graph-based Workflow Runtime (sequential/parallel/loop/conditional patterns with human-in-the-loop steps) or the Task API for structured agent-to-agent delegation.
3. 3Debug locally with the built-in ADK Web UI, and write repeatable evaluation test cases with adk eval to score trajectory and output quality before shipping.
4. 4Self-host the agent anywhere for free (Cloud Run, GKE, Docker, your own infra), or deploy to the managed Agent Runtime for autoscaling and built-in observability, billed as Google Cloud consumption.

Who is Google Agent Development Kit for?

• Teams that want Google's own official, actively-developed agent SDK rather than a third-party framework — especially anyone already on Gemini or Google Cloud
• Developers building multi-agent systems that need to interoperate across vendors, via native support for the Agent2Agent (A2A) protocol
• Polyglot teams — ADK is the only framework in this category shipping production-grade SDKs in four languages (Python, Go, Java, TypeScript) instead of one or two

Key features

• Workflow Runtime (ADK 2.0): A graph-based execution engine for composing deterministic, multi-agent flows — sequential, parallel, loop and conditional routing, with retry logic and human-in-the-loop steps.
• Task API: Structured, multi-turn delegation between agents — one agent can hand a task to another and track it to completion rather than a single-shot call.
• Native A2A + MCP protocol support: Built-in support for the Agent2Agent (A2A) protocol — now Linux Foundation-governed, with 150+ organizations reported in production use — for cross-vendor agent interop, plus native Model Context Protocol tool support.
• ADK Web + adk eval: A built-in local development/debugging UI (adk web) and a repeatable evaluation framework (adk eval) that scores an agent's full trajectory and final output, not just a spot-checked response.
• Four production SDKs + experimental Kotlin: Python (flagship, 2.0 GA), Go, Java and TypeScript all reached 1.0 GA, with an additional experimental Kotlin/Android runtime — the broadest language spread of any framework in this index.
• Agent Runtime (optional managed hosting): Deploy the same agent to Google Cloud's managed runtime (formerly Vertex AI Agent Engine) for autoscaling, session/memory persistence and built-in observability, without giving up the free self-hosted path.

What are Google Agent Development Kit's use cases?

• Cross-vendor multi-agent system: A team builds several specialist agents in ADK and has them delegate work to agents built on other frameworks or by other vendors over the A2A protocol, rather than being locked into one framework's internal orchestration.
• Polyglot organization standardizing on one agent SDK: A company with both a Python data-science team and a Go/Java backend team adopts ADK so every team builds agents against the same conceptual model and tooling, instead of maintaining separate frameworks per language.
• Self-hosted agent that can scale to a managed runtime later: A team starts by self-hosting an ADK agent on their own infrastructure for free, then moves the same code to the managed Agent Runtime once they need autoscaling and built-in observability without an infrastructure rewrite.

What does Google Agent Development Kit integrate with?

• Gemini, OpenAI, Anthropic Claude, Cohere, Ollama and 100+ providers via LiteLLM
• Model Context Protocol (MCP) tools, including MCP Toolbox for 30+ databases (Go SDK)
• Agent2Agent (A2A) protocol
• Vertex AI Search, Google Search grounding
• Google Cloud — Cloud Run, GKE, Vertex AI Agent Runtime
• Agent Designer (low-code visual builder, Gemini Enterprise Agent Platform)

Why use Google Agent Development Kit?

• Google's own official, actively-shipping agent SDK — Python alone pushed a new PyPI release (2.5.0) on 2026-07-16, and the Python repo saw a commit within the last 24 hours of this research.
• The broadest language spread of any framework in this index: production-grade Python, Go, Java and TypeScript SDKs, all reaching 1.0 GA, plus an experimental Kotlin/Android runtime.
• Native support for both dominant 2026 agent-interop standards (A2A and MCP) rather than bolting one on later — genuinely relevant if agents from different vendors need to talk to each other.
• Genuinely model-agnostic despite the Gemini optimization: OpenAI, Anthropic Claude, Cohere, local Ollama models and 100+ others are supported through its LiteLLM integration.
• Free, Apache-2.0, fully self-hostable core with an optional managed Agent Runtime layer for teams that want autoscaling without operating the infrastructure — the same 'free core, pay your own way' shape as this category's other major-lab frameworks.

Pros & cons

Google Agent Development Kit pricing

• Open source (self-host)Free
• Agent Runtime (managed, Google Cloud)$0.0864/vCPU-hour + $0.0090/GB-hour / usage-based

See current pricing on adk.dev ↗Compare Google Agent Development Kit alternatives →

Google Agent Development Kit review

Google Agent Development Kit is Google's answer to the same question Microsoft answered with Agent Framework and OpenAI answered with its own Agents SDK: a major AI lab's official, first-party toolkit for building agents, rather than leaving the whole category to third parties. What sets it apart is breadth and protocol alignment, not novelty — it's the only framework in this index shipping production-grade SDKs in four languages (Python, Go, Java, TypeScript, plus an experimental Kotlin/Android runtime), and it ships native support for both the Agent2Agent protocol and MCP rather than treating agent interoperability as an afterthought. It is genuinely model-agnostic despite the obvious Gemini optimization, and its free, Apache-2.0, fully self-hostable core pairs with an optional managed Agent Runtime for teams that want to scale without operating infrastructure themselves — the same shape Microsoft and OpenAI's official SDKs use. The honest catches are a real learning curve (this is a comprehensive framework, not a four-primitive minimal SDK), documentation and examples that still skew heavily toward Python even though the other three SDKs are GA, and a managed layer priced across several independent usage-based meters with no flat fee or hard spending cap. Pick it if you want Google's own official SDK, need genuine cross-vendor agent interoperability, or are standardizing on a language beyond Python; if none of those apply, LangChain's longer track record and larger ecosystem remain the safer default.

Google's official, protocol-forward agent SDK — the broadest language spread in this category (Python/Go/Java/TypeScript) with native A2A + MCP support, but still Python-first in its documentation and examples despite the other SDKs reaching GA.

Frequently asked questions