What is Google ADK (Agent Development Kit)?
Google's Agent Development Kit (ADK) is an open-source, Apache-2.0 framework for building, evaluating and deploying AI agents in Python, TypeScript, Go, Java and Kotlin, built for Gemini but able to run other models.
Checked
Of the three large model labs, Google is the one that explicitly recommends its own framework. Its Cloud Architecture Center points teams building agents on Google Cloud to ADK, and to Genkit for custom architectures. ADK agents get full integration on Agent Runtime, the hosting service that Google's docs now place under Gemini Enterprise Agent Platform.
ADK 2.0 changed what the framework is for. Version 1 was a toolkit for model-driven agents and fixed workflow agents. Version 2 added graph workflows that mix deterministic code with model reasoning, which puts ADK in the same space as LangGraph.
How it works
LLM agents
An agent built from instructions and tools that decides its own next step, and can grow into a team of agents that delegate to each other.
Workflow agents
Sequential, Parallel and Loop templates that run sub-agents in a fixed order.
Graph workflows (new in 2.0)
Nodes and explicit execution paths that combine deterministic code with model steps. Google's post on why it built ADK 2.0 (July 2026) reports that routing a refund flow through code cut tokens by about 50% and latency by about 20% in its example.
Models
Gemini first, with adapters for Gemma, Claude, OpenAI, Ollama, vLLM, LiteLLM and LiteRT-LM, including locally running models.
Tooling around the agent
A web Dev UI with a visual builder, an evaluation module with user and environment simulation, and A2A support for exposing and calling agents across frameworks.
Version and status, checked 1 November 2026
ADK 2.0 general availability
Python on 19 May 2026, Go on 30 June 2026 and TypeScript on 21 August 2026.
Latest releases
Python google-adk 2.13.0 (23 October 2026, Python 3.11 or later), TypeScript @google/adk 2.4.0 (28 October 2026), Go v2.6.0 (13 October 2026), Java v1.13.0 (21 October 2026) and Kotlin v1.5.0 (28 October 2026).
Adoption
22,067 GitHub stars on adk-python and 8,916 on adk-go, and 2.45M PyPI downloads in the week to 1 November 2026.
Deployment
Self-hosted, Cloud Run, GKE, or Agent Runtime. Agent Runtime also hosts LangChain, LangGraph, AG2 and LlamaIndex through managed templates, and other frameworks through a custom container.
When to pick it, and when not to
Pick it for
Teams already on Google Cloud, agents that run on Gemini, and organisations that need one framework across Python, TypeScript, Go and the JVM.
Also a fit when
Agents must call agents built in other frameworks over A2A, or a flow needs deterministic routing with model judgement only at specific nodes.
Check first when
The team writes Java or Kotlin. ADK 2.0 reached GA in Python, Go and TypeScript; the Java and Kotlin SDKs were on 1.x releases at this check, so graph workflow support should be confirmed per language.
Look elsewhere when
The stack sits outside Google Cloud and already runs on LangGraph, or the team wants a ready coding-style harness rather than a framework to build one.
Gotchas from production
2.0 is a breaking upgrade
In Python, BaseAgent now subclasses BaseNode, and overrides of _run_async_impl() are bypassed without an error. Custom agents need their logic moved into callbacks before upgrading.
Stored sessions change shape
2.0 events carry new fields such as node_info and output. Strict database schemas and JSON validators reject them, so stores shared between 1.x and 2.0 services need updating first.
Retries depend on exceptions
Broad except blocks inside tools swallow the errors the framework's retry policy needs. Google's guidance is to let exceptions propagate.
Deprecated workflow agents
In TypeScript, SequentialAgent, ParallelAgent and LoopAgent still run but log deprecation warnings. Go moves to the google.golang.org/adk/v2 import path.
Staying on 1.x is possible
Pinning google-adk~=1.0, @google/adk@^1.6.0 or the Go v1 module keeps an existing agent stable while a migration is planned.
How Motiontic approaches it
Motiontic's engineers are vendor-neutral. They run a short spike on the client's real task before committing to a framework or store, keep prompts, tool definitions and evals outside the framework so a later switch is a refactor rather than a rewrite, and hand the eval suite to the client's team so the next tooling decision can be measured rather than argued.
Google ADK and its nearest alternatives, checked 1 November 2026
| Google ADK | LangGraph | Microsoft Agent Framework | |
|---|---|---|---|
| Maintainer | LangChain Inc. | Microsoft | |
| Languages | Python, TypeScript, Go, Java, Kotlin | Python, JavaScript / TypeScript | Python, .NET (Go in a separate repo) |
| Licence | Apache-2.0 | MIT | MIT |
| Latest version (Python) | 2.13.0 (23 Oct 2026) | 1.2.19 (30 Oct 2026) | 1.24.0 (30 Oct 2026) |
| Workflow model | LLM agents, workflow agents, graph workflows | Explicit graphs with state and checkpoints | Graph workflows plus orchestration patterns |
| Default cloud | Google Cloud Agent Runtime | LangChain or Google managed runtimes | Microsoft Foundry |
| Agent-to-agent | A2A built in | A2A adopter | A2A, with 1.0 support announced as coming |
| PyPI downloads, week to 1 Nov 2026 | 2.45M | 13.60M | 396K (agent-framework-core) |
Common questions
Is Google ADK only for Gemini?
No. ADK is built for Gemini but ships adapters for Gemma, Claude, OpenAI, Ollama, vLLM, LiteLLM and LiteRT-LM, including locally running models.
What is new in ADK 2.0?
Graph-based, dynamic and collaborative workflows, which combine deterministic code with model reasoning along explicit execution paths. It also brought breaking changes to agent classes and event schemas.
Is Google ADK free?
The framework is open source under Apache-2.0. Costs come from the models it calls and from wherever the agent is hosted, such as Agent Runtime, Cloud Run or GKE.
What is the difference between ADK and Vertex AI Agent Engine?
ADK is the framework for building the agent. Google's hosting service for agents, formerly presented under Vertex AI, now appears in Google's docs as Agent Runtime on Gemini Enterprise Agent Platform.
Related
Google ADK vs LangGraph
Google's multi-language agent kit compared with LangChain's graph runtime.
Microsoft Agent Framework
Microsoft's one forward path for agents, replacing Semantic Kernel and AutoGen.
A2A protocol
The open protocol for agents talking to other agents, now under the Linux Foundation.
Agent graph
Nodes for model calls and code, edges for what runs next, and state that flows through.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Sources
- Google ADK docs (languages, agent types, models, deployment)
- Google ADK 2.0 (GA dates, breaking changes, migration)
- Google Developers Blog, ADK 2.0 design post (2026-07-01)
- GitHub, google/adk-python (22,067 stars, 1 Nov 2026)
- PyPI, google-adk (2.13.0, 23 Oct 2026)
- Google Cloud, Agent Runtime framework support (updated 2026-10-30)
- Google Cloud, Choose agentic AI architecture components
- Microsoft, Microsoft Agent Framework Version 1.0 (2026-04-03)
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.