Google Agent Development Kit: product and architecture
This Google Agent Development Kit review looks at ADK, the open agent framework Google publishes for building, evaluating, and deploying agents — and at the question that decides whether it is right for you: how much of your stack is already Google's. ADK is Apache-2.0 and free, with implementations in Python, TypeScript, Go, Java, and Kotlin, and a deployment path onto Vertex AI Agent Engine. We evaluated it against its documentation, its GitHub repositories, and the adoption signals we collect weekly.
Overview
ADK is Apache-2.0 licensed and published by Google. Our 14 September 2026 snapshot recorded about 21,500 GitHub stars on google/adk-python and roughly 2.4 million PyPI downloads a week for the google-adk package. A separate google/adk-java repository carries the Java implementation, and the project also ships TypeScript, Go, and Kotlin.
Its market position is the enterprise, polyglot end of the agent market. The project's own framing is "build production agents, not prototypes", and the five-language surface is the clearest expression of who it is for: organisations where the agent has to live inside an existing JVM or Go service rather than in a Python notebook. Version 2.0 introduced graph workflows, and the Python package was at v2.9.2 in mid-September 2026.
The audience is enterprise engineering teams, particularly those already on Google Cloud, who need agents that deploy into managed infrastructure and pass an evaluation gate before they do.
Key Features and Architecture
The base unit is an Agent with a name, a model, an instruction, and a list of tools — the same four fields in every language. From there, ADK's distinguishing features are structural rather than ergonomic.
Graph workflows, new in ADK 2.0, are the headline. They let you weave deterministic code with model reasoning in one structure: explicit execution paths where the process is known, model reasoning where it is not. This is the feature that separates ADK from frameworks that offer only a loop, and it is the reason a compliance-sensitive workflow can be expressed here without pretending the whole thing is deterministic.
Hierarchical multi-agent composition lets agents contain agents, so a coordinator delegates to specialists and the structure is a tree you can read rather than a set of handoff rules you have to trace.
The Agents CLI scaffolds, builds, tests, evaluates, and deploys from the terminal, and is designed to be driven from an AI-enabled editor.
Evaluation is first-party rather than an afterthought. ADK ships a harness for scoring agent behaviour, which matters because the thing that stops an agent reaching production is usually the absence of a pass/fail gate, not the absence of a feature.
Tooling includes Google's own — google_search is a one-line import — alongside an open ecosystem for connecting other services.
Deployment targets Vertex AI Agent Engine as a managed runtime, and this is where the Google gravity is strongest. ADK runs anywhere; the smooth path runs onto Google Cloud.
Ideal Use Cases
ADK earns its place in four situations. The first is a polyglot enterprise. If the agent has to be a Java service because the surrounding system is Java, ADK is one of very few frameworks that does not force a Python sidecar, and Go and Kotlin get the same treatment.
The second is a workflow that is partly deterministic. Graph workflows are the right shape when some steps must happen in a fixed order with explicit branches — an approval, a policy check, a write to a system of record — and only some steps need a model to decide.
The third is a team on Google Cloud that wants a managed runtime. Deploying to Vertex AI Agent Engine removes the operational question entirely, and the integration is first-party rather than reconstructed.
The fourth is any team that has to demonstrate an agent works before shipping it. The built-in evaluation harness is the mechanism, and adding one to a framework that lacks it is real work.
Do not choose ADK if you are a small Python team shipping one agent. The five-language surface, the graph model, and the deployment tooling are all answers to enterprise problems, and you will pay for them in concepts you do not need — PydanticAI or the OpenAI Agents SDK will get you there faster. It is also the wrong pick if you are committed to another cloud and want the managed path: ADK will run on AWS or Azure, but Agent Engine is the part that makes the deployment story compelling, and that part is Google's.
Pricing and Licensing
ADK is free under the Apache 2.0 licence, with no paid tier of the framework, no seat charge, and no enterprise edition holding features back. Apache 2.0 is permissive enough for commercial use and redistribution, and it satisfies an open-source-only licence requirement — which is worth noting because it applies to the whole framework, including the graph workflows and the evaluation harness, not just a community core.
The costs sit in two places. The first is model spend, billed by whichever provider the agent calls. The examples use Gemini, and if you use Gemini, that is Google's per-token bill; ADK does not require it, and pointing an agent at another provider is a model string change.
The second is infrastructure, and it is the one to plan for. Running ADK yourself costs whatever your compute costs. Deploying to Vertex AI Agent Engine is a managed Google Cloud service with its own usage-based charges, separate from the model tokens it consumes. That is the honest comparison against a self-hosted framework: ADK's software is free at both ends, but the deployment path the product is designed around is a paid service.
Compared with the alternatives, the licensing is the most permissive position in the category alongside the OpenAI Agents SDK, PydanticAI, and LangGraph, all of which are free with no paid tier. CrewAI and LangChain are Freemium, with a paid platform beside the open framework. Budget for the model calls and, if you use it, for Agent Engine.
Strengths & Trade-offs
Pros
- Five first-party language implementations — Python, TypeScript, Go, Java, Kotlin — which is rare and decisive for a JVM or Go shop.
- Apache 2.0 across the whole framework, so no feature sits behind a plan and an open-source-only requirement is satisfied.
- Graph workflows express a process that is partly deterministic, instead of forcing every step through a model.
- A first-party evaluation harness, so "does this agent work" has an answer before deployment rather than after.
- A managed deployment target in Vertex AI Agent Engine, which removes the runtime question for Google Cloud teams.
Cons
- The centre of gravity is Google Cloud. The framework is portable; the deployment story that makes it attractive is not.
- More concepts than a small team needs. Hierarchical composition, graph workflows, and the CLI are enterprise answers, and they cost learning time even when unused.
- Version 2.x is recent, and the 2.0 graph-workflow redesign means older tutorials and examples describe a different framework.
- The five language implementations do not all move together, so a feature announced for Python may not be in the Go or Kotlin build yet.
- Documentation assumes Gemini in most examples, so using another provider means translating as you read.