Google ADK
by Google
Google's open-source, code-first framework for building and deploying production AI agents.
Google's Agent Development Kit is an open-source, code-first framework for building, evaluating and deploying AI agents, with SDKs for Python, Java, Go and TypeScript. It is model-agnostic — running on Gemini, Claude, OpenAI models or local runtimes — and pairs graph-based workflows with managed context so multi-agent systems stay debuggable in production rather than only in demos.
Google's Agent Development Kit (ADK) is an open-source, code-first framework for building, evaluating and deploying AI agents, released in April 2025 and now shipping as ADK 2.0 GA. It is model-agnostic by design: agents run on Gemini and Gemma but also Anthropic Claude, OpenAI models and locally hosted runtimes such as Ollama and vLLM, so the framework is not a wrapper around a single API. ADK 2.0's headline addition is graph workflows, which let developers interleave deterministic code with model-directed reasoning instead of relying purely on LLM delegation — the specific gap that had pushed teams toward LangGraph when they needed predictable routing. Its other distinguishing decision is explicit context management: ADK treats agent context as a managed artefact rather than concatenated strings, which is what makes long-running multi-agent sessions debuggable. The kit ships multi-agent orchestration primitives, custom and OpenAPI tools, Model Context Protocol support, native agent-to-agent (A2A) protocol support, a local web UI and CLI for debugging, and an evaluation harness. SDKs exist for Python, Java, Go and TypeScript/JavaScript, with community ports including Rust. Deployment targets include Google Cloud's Agent Runtime, Cloud Run and GKE, or any container platform. The Python repository is Apache-2.0 licensed and has accumulated roughly 21,000 GitHub stars and 3,800 forks since launch, with commits continuing through August 2026 and around 550 open issues. The framework itself is free; cost comes from the underlying model calls and whatever infrastructure hosts the agents, the usual pattern for vendor-sponsored open-source agent frameworks.
Platform engineering teams already standardised on Google Cloud who need multi-agent systems in production with deterministic routing and an evaluation story, not a prototype.
Graph workflows and managed context make agent behaviour reproducible and debuggable, so an agent that works in testing behaves the same way under load.
At a Glance
- Category
- Agent Development
- Pricing
- Free, Usage-based
- Target Market
- CTOs, Platform Engineers, Enterprise Developers, ML Engineers, AI Architects
- Deployment
- Open-source, Self-hosted, Cloud-first, Hybrid
- Founded
- 2025
- Headquarters
- Mountain View, United States
- Team Size
- 500+
Key Features
- ✓Graph workflows
Weaves deterministic code with adaptive model reasoning so routing is explicit and reproducible rather than left entirely to LLM delegation.
- ✓Managed context
Treats agent context as a first-class managed artefact instead of concatenated strings, which is what keeps long multi-agent sessions debuggable.
- ✓Model-agnostic execution
Runs on Gemini, Gemma, Anthropic Claude, OpenAI models and local runtimes such as Ollama and vLLM without rewriting the agent.
- ✓Multi-agent orchestration
First-class primitives for agent teams and collaborative workflows, with native agent-to-agent (A2A) protocol support for cross-system communication.
- ✓Tooling and MCP support
Custom tools, OpenAPI-defined tools and Model Context Protocol servers, so existing APIs become agent-callable without bespoke adapters.
- ✓Built-in evaluation and debugging
A local web UI, CLI and evaluation harness let teams inspect and score agent trajectories before promoting anything to production.
- ✓Multi-language SDKs
Python, Java, Go and TypeScript/JavaScript implementations let a polyglot organisation standardise on one agent architecture.
Capabilities
Use Cases
- •Production customer-support agent teams
A supervisor agent routes to specialist sub-agents with deterministic escalation paths that survive audit and load testing.
- •Internal operations automation on Google Cloud
Agents call existing internal APIs through OpenAPI tools and deploy straight to Cloud Run or GKE.
- •SRE and infrastructure assistants
Community projects use ADK for Kubernetes and AWS operations agents that diagnose incidents and propose remediation steps.
- •Cross-vendor agent interoperability
A2A protocol support lets agents built on different stacks exchange work without a bespoke integration for each pair.
- •Evaluated agent rollout
Teams score agent trajectories against a fixed evaluation set before promoting a new prompt or model version to production.
Ideal For
Best For
- ✓Multi-agent systems on Google Cloud where Agent Runtime, Cloud Run or GKE is already the deployment target
- ✓Workflows that must interleave deterministic business logic with model reasoning rather than delegating routing entirely to an LLM
- ✓Teams standardising on open protocols — MCP for tools and A2A for agent-to-agent communication
- ✓Polyglot organisations needing the same agent framework in Python, Java, Go and TypeScript
- ✓Enterprises that require an open-source, self-hostable framework rather than a proprietary agent SaaS
Not Ideal For
- ✗Teams outside the Google Cloud ecosystem — reviewers consistently flag that the tightest integration is with Gemini and Google Cloud, and other providers require more wiring through API wrappers
- ✗Anyone needing the deepest ecosystem of prebuilt integrations and community examples today, where LangGraph remains further ahead on adoption and third-party material
- ✗Non-developers looking for a visual agent builder — ADK is deliberately code-first with no drag-and-drop authoring surface
- ✗Projects that cannot absorb framework churn: SDKs outside Python have carried beta status, and a 2.0 release this recent will keep moving
Integrations
Deployment
Market Analysis
Pros
- ✓Apache-2.0 and self-hostable, so there is no licence cost and no hard lock-in to Google Cloud
- ✓Model-agnostic across Gemini, Claude, OpenAI and local runtimes such as Ollama and vLLM
- ✓Graph workflows in 2.0 close the determinism gap reviewers previously cited as the reason to pick LangGraph
- ✓Backed by Google with active development — roughly 21,000 stars, 3,800 forks and commits continuing through August 2026
- ✓Ships evaluation and debugging tooling in the core, which most agent frameworks leave to third parties
Cons
- ✗Reviewers consistently report Google Cloud gravity — the tightest integration is with Gemini and Google Cloud, and other providers need more work through API wrappers
- ✗SDKs outside Python have carried beta status, so the multi-language story is less mature than the language list suggests
- ✗The ecosystem of third-party integrations, tutorials and community examples still trails LangGraph, which holds a large adoption lead
- ✗Around 550 open issues on the Python repository indicates a framework still absorbing rapid change
- ✗A 2.0 release this recent means API churn is likely, which is a real cost for teams with agents already in production
Pricing
Open source (Apache-2.0)
$0
- ✓Full framework
- ✓All SDKs
- ✓Self-hosted anywhere
- ✓No licence fee
Managed on Google Cloud
Contact for pricing
- ✓Agent Runtime
- ✓Cloud Run / GKE deployment
- ✓Vertex AI model hosting
- ✓Billed as standard Google Cloud consumption
The framework itself is free and Apache-2.0 licensed, with no seat, node or licence charge — the entire cost is downstream. Real spend comes from model inference (Gemini, Claude, OpenAI or whatever you point it at) plus the infrastructure running the agents, whether that is Google Cloud Agent Runtime, Cloud Run, GKE or your own containers. There is no published list price for ADK because there is nothing to price; budget against Google Cloud consumption and per-token model billing instead, and note that self-hosting on non-Google infrastructure is fully supported and removes the cloud line entirely.
Security & Compliance
Connect
Sources
This page was written from 6 sources, 5 on domains other than adk.dev.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Mastra
Open-source TypeScript framework and platform for building, deploying and observing production AI agents and workflows
Agentrys
Agentic design automation that lets chip teams build and own a self-improving engineering workforce
OpenAI Agents API
The Codex agent harness behind one API call — managed sessions, sandboxes and subagents
Letta
Open-source platform for stateful AI agents with persistent memory, now with a TypeScript Agents SDK for embedding them in your own apps
Mentioned In
LangGraph vs Google ADK: Which Enterprise AI Framework Wins?
Side-by-side comparison of LangChain's LangGraph and Google's Agent Development Kit for enterprise AI development. Real feature analysis, deployment costs, and use case recommendations for technical leaders and engineering leaders.
March 22, 2026MCPMCP vs LangChain vs OpenAI Functions: Which for Enterprise?
Choosing between MCP, LangChain Tools, and OpenAI Functions isn't an either/or decision—many teams use MCP for standardized data access alongside LangChain for orchestration. The real question is which to prioritize for your enterprise use case.
March 22, 2026GeelyGeely Expands NVIDIA Partnership Across Physical, Enterprise, and Industrial AI
Geely Auto Group announced a major expansion of its NVIDIA partnership spanning autonomous vehicles, cloud-based agentic AI, and factory automation—signaling a transformation from car manufacturer to AI organization.
March 22, 2026Cloud InfrastructureMicrosoft Loses OpenAI Exclusivity as AWS Pays $50B: What Enterprise AI Buyers Should Do Now
Microsoft Loses OpenAI Exclusivity as AWS Pays $50B. For enterprise decision-makers: strategic analysis, cost implications, and implementation guidance for A...
March 22, 2026Enterprise AIOpenAI Turns ChatGPT Into a Sales Channel for Retailers
OpenAI launched Agentic Commerce Protocol expansion for product discovery — Target, Walmart, and Shopify are integrated. High-intent shoppers meet conversational AI.
March 25, 2026