Google ADK
Google ADK is an Apache 2.0 open-source framework for building, evaluating, and deploying multi-agent systems, with first-class support for Gemini and Google Cloud and coverage across Python, TypeScript, Go, Java, and Kotlin.
What is Google ADK?
Google ADK is an Apache 2.0 open-source framework introduced at Google Cloud NEXT 2025 for developing, debugging, evaluating, and deploying AI agents and multi-agent systems. It is the same framework powering agents in Google products such as Agentspace and the Google Customer Engagement Suite. ADK is designed to cover the entire agent development lifecycle rather than only the orchestration layer: it ships a local web UI for inspecting events and state, built-in evaluation tooling that scores both final responses and execution trajectories, and deployment scaffolding for Cloud Run, Gemini Enterprise Agent Runtime, and Google Kubernetes Engine.
ADK 2.0, generally available for Python as of May 2026 and for Go and TypeScript shortly after, shifted the core execution model from a hierarchical agent executor to a graph-based Workflow Runtime. Agents, tools, and functions are now evaluated as nodes within a workflow graph, with support for routing, fan-out/fan-in, loops, retries, state management, human-in-the-loop pauses, and nested workflows. The framework supports five languages—Python, TypeScript, Go, Java, and Kotlin—each built for its own runtime and type system while sharing open protocols for interoperability, including MCP for tools and A2A for cross-framework agent communication.
How does Google ADK work?
A developer defines agents as code, specifying instructions, tools, and model configuration. A Workflow Runtime executes the graph, routing between agents and tools according to deterministic edges or LLM-driven transfer logic. Multi-agent hierarchies allow a coordinator agent to delegate to specialist sub-agents; those agents can run sequentially, in parallel, in loops, or in nested workflows, with structured results returned to the parent.
The framework provides a rich tool layer: pre-built tools such as Google Search and code execution, MCP servers, OpenAPI-spec tools, third-party libraries (LangChain, LlamaIndex), and other agents or frameworks (LangGraph, CrewAI) wrapped as tools. A bidirectional audio and video streaming capability supports real-time multimodal interaction beyond text. Evaluation is built into the development loop: developers can compare agent trajectories and final outputs against test cases without leaving the ADK toolchain. Tracing follows OpenTelemetry GenAI conventions, so spans can be exported to Cloud Trace, Jaeger, Datadog, or Langfuse.
Deployment is container-based: agents can run anywhere, with first-class paths to Cloud Run and Google Cloud's Agent Runtime for managed sessions and memory. The adk web dev UI provides a local browser interface for chatting with agents, switching sessions, inspecting state and events, and—in Python—sketching agents in a visual builder. That UI is explicitly a development tool, not a production control plane.
Use Cases
- Building production multi-agent systems on Google Cloud that require managed hosting, evaluation, and tracing within the Gemini ecosystem
- Composing hierarchical agent teams where a coordinator delegates to specialists, with deterministic workflow control over routing and retries
- Creating real-time voice or video agents using ADK's bidirectional streaming, without assembling a separate media pipeline
- Standardising agent development across polyglot engineering organisations that need the same architectural model in Python, Go, Java, Kotlin, and TypeScript
- Prototyping multi-agent applications with a local UI and built-in evaluation before committing to a deployment target
Pricing
Pricing model: Freemium
| Plan | Price | Billing | Included usage | Notes |
|---|---|---|---|---|
| Open Source Framework | Free | — | Apache 2.0 licensed in Python, TypeScript, Go, Java, and Kotlin, including the adk web dev UI, evaluation tools, and API server. No framework licensing fee. Costs come from the underlying model API calls and any cloud infrastructure used. | — |
| Agent Runtime (Google Cloud) | Usage-based | per vCPU-hour and GiB-hour | Managed hosting for ADK agents on Google Cloud with managed sessions and memory. Billed at approximately $0.085 per vCPU-hour and $0.009 per GiB-hour above a no-cost tier of 50 vCPU-hours and 100 GiB-hours per account per month. A Google Cloud project with billing enabled is required. | — |
Common Limitations
- Gemini and Google Cloud orientation · ADK is optimized for Gemini models and Google Cloud deployment. Other models are supported through LiteLLM or dedicated connectors, but the framework's defaults, documentation, and deployment paths are Gemini-first. Teams not using Gemini or Google Cloud should expect additional integration work relative to a provider-neutral stack.
- No built-in enterprise governance · The framework does not include SSO, role-based access control, or audit logs. These must be added through surrounding infrastructure or organisational processes. The hosted Agent Runtime provides managed hosting but does not automatically supply identity or permission layers.
- Language-specific feature lag · Versions differ by language, and features land in Python first. The visual builder in the adk web dev UI is Python-only. Teams using Go, Java, Kotlin, or TypeScript may need to wait for parity on newer capabilities, or to implement missing functionality themselves.
- Concurrency and observability rough edges · Public issue reports describe intermittent MCPToolset failures under concurrent multi-agent load, parent-agent call_llm spans not being exported to OpenTelemetry backends in transfer_to_agent setups, and parallel sub-agent branches causing downstream mergers to see only one branch's output. These are active areas of fix activity and may not affect all configurations, but teams running high-concurrency or deeply nested multi-agent systems should validate tracing and tool execution under realistic load.
- Unbounded execution risk · Agent.Run() delegates to the ADK runner without a built-in turn cap. A stuck tool-call chain can spin indefinitely. Teams need to implement their own guardrails or monitor runtime behaviour; the framework does not provide a default ceiling on loop iterations or tool invocations.
Key Features
- Graph-based Workflow Runtime (ADK 2.0)
- Hierarchical multi-agent delegation
- Python, TypeScript, Go, Java, and Kotlin support
- MCP and A2A protocol interoperability
- Local adk web dev UI with event inspection
Google ADK vs. LangGraph and CrewAI
This comparison uses current official documentation and vendor-published material. It focuses on documented architecture, deployment model, language coverage, and governance scope, not subjective rankings of output quality or speed.
| Tool | Positioning | Key differences | Best fit | Pricing note |
|---|---|---|---|---|
| Google ADK Official ADK site Agent Runtime pricing | Google's open-source, code-first framework for the full agent lifecycle, with a graph-based Workflow Runtime and deep Gemini and Google Cloud integration, available in five languages. | Covers build, evaluate, and deploy in one framework, with a local dev UI, built-in trajectory evaluation, MCP and A2A support, and managed hosting on Agent Runtime. Optimized for Gemini and Google Cloud; other models go through LiteLLM or connectors. No built-in SSO, RBAC, or audit logs. | Teams building production multi-agent systems on Google Cloud or standardising agent development across Python, Go, Java, Kotlin, and TypeScript. | Apache 2.0, free framework. Agent Runtime usage-based at ~$0.085/vCPU-h and ~$0.009/GiB-h above 50 vCPU-h and 100 GiB-h monthly no-cost tier. |
| LangGraph LangGraph overview LangGraph Platform plans | A low-level orchestration framework and runtime for stateful, long-running agents modelled as explicit graphs of shared state, nodes, and edges. | Prioritises explicit graph control with durable checkpointing, human-in-the-loop interrupts, and time-travel debugging. Does not prescribe an agent architecture or prompt structure. Managed deployment and observability are available through LangSmith. Python and JavaScript/TypeScript primary. | Teams that need fine-grained control over state transitions and a mature production runtime with checkpointing, independent of a single cloud vendor. | MIT-licensed SDK. LangSmith Developer is free; Plus is $39 per seat per month plus usage-based deployment pricing. Enterprise custom-quoted. |
| CrewAI CrewAI introduction CrewAI pricing | An open-source framework for role-playing agent teams (Crews) within structured, stateful Flows, with a hosted enterprise platform for governance and deployment. | Combines a role-based Crew abstraction with deterministic Flows. The open-source core is MIT-licensed; built-in tracing, SSO, RBAC, and audit trails live in the hosted AMP product. Ships 40+ pre-built tools. Python-primary. | Teams that want a readable role-and-task abstraction for multi-agent pipelines and an optional governed hosted runtime without adopting a graph-first programming model. | Open source free under MIT. AMP Basic free with 50 executions/month; Enterprise custom-quoted. |
Comparison checked against official product information on 2026-11-05. Features and pricing may change.
Frequently Asked Questions
Is Google ADK free?
Yes. The ADK framework is Apache 2.0 licensed and free to use in all five supported languages, including the adk web dev UI, evaluation tools, and API server. Google Cloud's Agent Runtime is billed separately by vCPU and memory hours above a monthly no-cost tier. Model API costs are also separate.
Does Google ADK only work with Gemini?
No, but it is optimized for Gemini. The framework supports other models through LiteLLM integration and dedicated connectors, including Anthropic, Meta, Mistral, and models accessible via Vertex AI Model Garden. Teams using non-Gemini models should expect additional configuration and may lose some Gemini-specific conveniences.
What changed in ADK 2.0?
ADK 2.0 replaced the hierarchical agent executor with a graph-based Workflow Runtime. Agents and tools are now evaluated as nodes in a workflow graph, enabling more explicit control over routing, retries, fan-out, loops, and state. The release introduced breaking changes for custom session storage schemas, execution-driver overrides, and error handling; official migration guidance is available.
Does ADK include a visual builder?
The adk web dev UI includes a visual builder, but it is Python-only and intended for development, not production. Go, Java, Kotlin, and TypeScript users have the CLI and web UI for inspection but not the visual agent-sketching capability.
How does Google ADK handle enterprise security and access control?
Not in the framework itself. ADK does not include SSO, role-based access control, or audit logs. These need to be layered in through the surrounding deployment environment. The hosted Agent Runtime provides managed hosting but does not supply identity or permission management as part of the ADK framework.
Alternative Tools
-
OpenAI Agents SDK OpenAI Agents SDK is a lightweight, MIT-licensed Python and TypeScript framework for building multi-agent workflows using a minimal set of primitives: agents, tools, handoffs, guardrails, sessions, and tracing. -
Vercel AI SDK Vercel AI SDK is a free, Apache 2.0 TypeScript toolkit for building streaming AI UIs and agents with React, Next.js, Vue, Svelte, and Node.js. It unifies 100+ models across providers behind a single API, ships React hooks for chat and generative UI, and requires Node.js 22+ with ESM-only imports in v7. It is not a RAG framework or a full agent orchestrator. -
CrewAI CrewAI is an open-source Python framework for orchestrating role-playing AI agent teams (Crews) within structured, stateful workflows (Flows), with a hosted AMP platform for enterprise governance. -
Dify Dify is an open-source platform for building, deploying, and operating LLM applications, agents, RAG pipelines, and workflows through a visual canvas, with cloud-hosted and self-hosted deployment options. -
LangChain LangChain is an MIT-licensed framework for building agents and LLM-powered applications, with create_agent as its current entry point and LangGraph as the underlying orchestration runtime. LangSmith, a separate paid product, adds tracing, evaluation, and deployment. -
Browser Use Browser Use is an MIT-licensed open-source AI browser agent (Python and TypeScript) that lets LLMs control a browser, paired with a usage-based cloud offering at $0.02 per browser-hour for managed stealth browsers, CAPTCHA solving, and a hosted agent API.