Google Agent Development Kit (ADK) is an open source framework from Google for building, debugging and deploying AI agents, announced on 9 April 2025 at Google Cloud Next 2025 and now at version 2 in Python, Go and TypeScript. Google's documentation calls it "the open-source agent development framework that lets you build, debug, and deploy reliable AI agents". People meet it under several names: the ADK libraries themselves, the adk.dev documentation site (the old google.github.io/adk-docs address now redirects there), and the managed Google Cloud services it deploys to. This page sets out which languages ADK supports, what changed in ADK 2.0, how agents are built and deployed, which models it works with, what it costs and how the release cadence works. It is written for developers, architects and the people who approve frameworks for a team. Every fact comes from Google's ADK documentation, the ADK repositories and Google Cloud's own pages. Last checked: 2026-11-15.
This is an independent guide from NHI Mgmt Group, based on Google's public ADK documentation as of November 2026. It is not sponsored or reviewed by Google. Product features change, so confirm settings against Google's current documentation.
Key takeaways
- Five languages, one framework. ADK is published for Python, TypeScript, Go, Java and Kotlin, each in its own Apache 2.0 repository under the google GitHub organisation.
- ADK 2.0 is a new runtime, not a point release. Python 2.0 reached GA on 19 May 2026, Go on 30 June 2026 and TypeScript on 21 August 2026, moving from a hierarchical agent executor to a graph-based Workflow Runtime with breaking changes to events and sessions.
- The framework is free; hosting and models are not. ADK itself costs nothing. Spend comes from the model you call and from where the agent runs, such as Agent Runtime on Gemini Enterprise Agent Platform, Cloud Run or GKE.
- Model choice is open. ADK resolves Gemini, Claude and Agent Platform hosted models by name, and connects to others through connector classes such as
LiteLlmandApigeeLlm, plus Ollama, vLLM and LiteRT-LM. - Python moves fastest. The Python README states a release cadence of "roughly bi-weekly", and the Python repository showed v2.14.1 on 12 November 2026 with a v1.39.1 bug-fix backport in August 2026.
What is Google Agent Development Kit (ADK)?
ADK is a code-first framework: you define agents, tools and orchestration in ordinary Python, TypeScript, Go, Java or Kotlin, then run them locally or deploy them. Google's launch post described it as "the same framework powering agents within Google products like Agentspace" and the Google Customer Engagement Suite. Google says ADK is optimised for Google Cloud but works with "almost any generative AI model" and can be containerised to run on any infrastructure.
The documentation groups ADK's building blocks as follows:
- Agents: the
LlmAgent(also calledAgent), a model-powered unit with instructions and optional tools; template workflow agents (SequentialAgent,ParallelAgent,LoopAgent); and custom agents. - Graph workflows: new in ADK 2.0, composable execution paths that combine model reasoning with deterministic code, where agents, tools and functions run as individual nodes.
- Core concepts: sessions, state, memory, artifacts, callbacks and events.
- Tools: custom function tools, OpenAPI tools, MCP tools and remote agents over the A2A protocol, plus a catalogue of integrations.
- Context management: the site says ADK filters irrelevant events, summarises older turns, lazy-loads artifacts and tracks token usage.
Every Python ADK project needs one thing: agent.py must define a root_agent. Tools are optional.
| Building block | What it is for | Typical use |
|---|---|---|
LlmAgent | Model reasoning with instructions and tools | A single assistant or a specialist sub-agent |
| Template workflow agents | Fixed sequential, parallel or loop execution | Pipelines built before 2.0 (deprecated in TypeScript 2.0 in favour of graph workflows) |
| Graph workflows | Routing, fan-out and fan-in, loops, retries, human-in-the-loop, nested workflows | Processes that mix deterministic steps with model decisions |
| Callbacks | Hooks before and after agent, model and tool calls | Per-agent validation, such as checking tool parameters |
| Plugins | Policies applied once at the runner level | Organisation-wide guardrails across every agent |
| Agent Config | Agents defined without code | Simple agents maintained by configuration |
Key facts at a glance
As of 15 November 2026, from adk.dev, the ADK repositories and the Google Developers Blog.
| Field | Value |
|---|---|
| Owner | |
| Type | Open source, code-first agent development framework |
| Announced | 9 April 2025, Google Cloud Next 2025 (Google Developers Blog) |
| Languages | Python, TypeScript, Go, Java, Kotlin |
| Licence | Apache 2.0 (adk-python, adk-js, adk-go, adk-java and adk-kotlin READMEs) |
| Current major version | 2.x for Python, Go and TypeScript; Java README install snippet shows 1.14.0; Kotlin README shows 1.5.0 |
| Latest Python release seen | v2.14.1, 12 November 2026 (repository CHANGELOG) |
| Python requirement | 3.11 or later; constraints files for 3.11 to 3.15 |
| Node.js requirement | 20.19 or newer (adk-js README) |
| Kotlin requirement | Gradle or Maven; JDK 21 or later for the litertlm module |
| Python release cadence | "Roughly bi-weekly" (adk-python README) |
| Documentation | adk.dev (google.github.io/adk-docs redirects there) |
| Price | Free and open source; model and hosting usage billed by the provider you use |
Plans and pricing
ADK has no plans or seats: the libraries are Apache 2.0 and free to use. Costs come from two places.
- Models: calls to Gemini through the Gemini API or Agent Platform, or to any other provider through a connector, are billed by that provider. The Python quickstart uses a
GOOGLE_API_KEYfrom Google AI Studio stored in the agent's.envfile. - Hosting: deployment targets bill for their own resources. On Gemini Enterprise Agent Platform, Agent Runtime is metered on Agent Compute and Agent Memory.
The Agent Platform pricing page, as of 15 November 2026, lists these list prices in US dollars for agent hosting:
| Resource | List price | Notes from the pricing page |
|---|---|---|
| Agent Compute | $0.085 per vCPU-hour | 1-year and 3-year consumption rates are lower ($0.0765 and $0.068) |
| Agent Memory | $0.009 per GiB-hour | Usage rounded to the nearest second |
| Idle time between turns | Not billed | Time waiting for the next prompt is excluded |
| Sessions and Memory Bank storage | $0.30 per GiB-month | Billing started 1 September 2026; reads and writes metered in vCPU-hours |
The Agent Platform product page also lists a $300 free trial credit for new Google Cloud customers, valid for 90 days, and a monthly free tier per billing account for agent runtimes (first 50 hours of compute, 100 hours of RAM and 1 GiB of storage). Cloud Run, GKE and other container platforms bill under their own pricing. See the Gemini Enterprise Agent Platform page for the managed platform in full.
Models available
ADK does not ship a model. The models page describes three ways to connect one, and states support for Python, TypeScript, Go and Java:
| Mechanism | How it works | Models named in the docs |
|---|---|---|
| Direct string (registry) | Pass a model name or endpoint string; ADK's registry resolves the backend | Claude, Gemini, Agent Platform hosted models |
| Model connectors | Instantiate a wrapper class and pass it as the model of an LlmAgent | Apigee (ApigeeLlm), LiteLLM (LiteLlm), LiteRT-LM, Ollama, vLLM |
| Model routing | A router function picks a model at runtime and fails over automatically on error | Any configured models |
The documentation navigation also lists Gemma and OpenAI model pages. The Python quickstart's example agent uses gemini-flash-latest. Kotlin's README lists Anthropic and OpenAI integrations on the JVM, and Firebase AI, ML Kit and on-device LiteRT-LM on Android. Locally running models are supported through adapters such as Ollama and vLLM, which suits offline or disconnected deployments.
Features and modes
ADK 2.0 Workflow Runtime: the 2.0 overview says the runtime moves from a hierarchical agent executor to a graph-based engine. It adds three styles of workflow:
- Graph-based workflows: deterministic routing and execution control.
- Dynamic workflows: code-based logic for loops and complex branching.
- Collaborative workflows: coordinator agents working with sub-agents.
The Python README adds a Task API for structured agent-to-agent delegation in multi-turn and single-turn modes, and Tool Confirmation, which asks a human to approve before a tool runs. Go v2.0.0 added a workflow package, agent execution modes and human-in-the-loop tool confirmation.
Tools and integrations: the integrations catalogue is filtered into Code, Connectors, Data, Evaluation, Google, MCP, Observability, Resilience and Search. Google tools include Google Search, Code Execution (Gemini), Computer Use, BigQuery, Bigtable and Spanner tools, Cloud Storage, Pub/Sub, Google Apps toolsets and Secret Manager. Code execution options include the Code Execution Tool with Agent Runtime and a GKE Code Executor. MCP Toolbox for Databases and several vector databases sit under data. The TypeScript README lists built-in Google Search, Google Maps, Vertex AI Search and URL context tools, and notes that ADK TypeScript runs on servers and in browsers.
Observability: Cloud Trace is the Google option; the catalogue also lists third-party tracing integrations such as Arize AX, AgentOps, Datadog, Langfuse, MLflow Tracing and Phoenix. Kotlin exports OpenTelemetry traces over OTLP.
Safety features: the safety page documents three plugins that apply at runner level: a Gemini as a Judge plugin that screens inputs, tool calls and responses for unsafe content, prompt injection and jailbreaks; a Model Armor plugin that ships with ADK; and a PII Redaction plugin that redacts personal data in the Before Tool Callback.
Supported IDEs and platforms
ADK runs wherever its language runtime runs. The install commands on adk.dev are:
| Language | Package or module | Install |
|---|---|---|
| Go | google.golang.org/adk/v2 | go get google.golang.org/adk/v2 |
| Java | com.google.adk:google-adk (dev UI: google-adk-dev) | Maven or Gradle dependency |
| Kotlin | com.google.adk:google-adk-kotlin-core | Maven Central; -jvm artifact for Maven JVM projects |
| Python | google-adk (optional google-adk[extensions]) | pip install google-adk |
| TypeScript | @google/adk (CLI and dev UI: @google/adk-devtools) | npm install @google/adk |
Languages are listed alphabetically. For local development the Python CLI provides adk create (scaffolds agent.py, .env and __init__.py), adk run (interactive command line) and adk web (a browser chat UI on a local port such as 8000). The quickstart states that ADK Web is "not meant for use in production deployments". Kotlin offers AdkApiServer (headless HTTP) and AdkDevServer (with a development UI). Google also provides an Agents CLI to scaffold, test, evaluate and deploy agents, and ADK developer Skills for AI-assisted coding.
The deployment page lists four targets:
| Target | What Google's docs say it is | When it fits |
|---|---|---|
| Agent Runtime (Gemini Enterprise Agent Platform) | Fully managed, auto-scaling service for deploying and scaling agents | You want the most managed option, with sessions and memory as services |
| Cloud Run | Managed, auto-scaling container platform | You want a container you control with serverless scaling |
| Google Kubernetes Engine (GKE) | Managed Kubernetes | You need more deployment control or want to serve open models alongside |
| Other container infrastructure | Any environment that runs container images, such as Docker or Podman | On-premises, offline or disconnected setups |
Earlier ADK material, including the 2025 launch post and the Python README, names the managed target "Vertex AI Agent Engine"; current adk.dev pages call it Agent Runtime on Agent Platform.
Enterprise and admin controls
ADK is a library, so admin controls come from where it runs. The safety page describes two identity patterns for tools: Agent-Auth, where the tool acts with the agent's own identity such as a service account (suitable when all users share the same access, with user attribution logged by you), and User Auth, where the tool uses the controlling user's identity through OAuth so the agent can only do what that user could. It also recommends running agents inside a VPC Service Controls perimeter to keep API calls within it, paired with tool-level guardrails because perimeters are coarse. On Agent Platform, agents can additionally use Agent Identity, Agent Registry and Agent Gateway, covered on the Gemini Enterprise Agent Platform page. See the Google Cloud Agent Identity guide and the VPC Service Controls Guide for the identity and perimeter detail.
Data handling and privacy
What leaves your environment depends on configuration rather than on the framework:
| Component | Where data goes | Control |
|---|---|---|
| Model calls | The model provider you configure (Gemini API, Agent Platform, or another provider via a connector) | Choose the provider and endpoint; local models through Ollama, vLLM or LiteRT-LM keep inference on your infrastructure |
| Sessions, state and memory | Your chosen session service or database; Agent Runtime Sessions and Memory Bank when hosted on Agent Platform | Pick the storage backend; 2.0 sessions add new event fields that rigid schemas must store |
| Tool calls | Whichever APIs or MCP servers the tools reach | In-tool guardrails, callbacks and the PII Redaction plugin |
| Traces | Cloud Trace or a third-party observability integration you enable | Enable only the exporters you need |
| Code execution | A sandbox such as the Code Execution Tool with Agent Runtime or the GKE Code Executor | Google's safety page says custom executors should have no network access and clear data between runs |
Credentials for local development live in the agent's .env file in the quickstart, and the Kotlin README uses Application Default Credentials with a GOOGLE_CLOUD_PROJECT value for Vertex-backed Gemini. The safety page also warns that model-generated content must be escaped in user interfaces, because unescaped HTML or JavaScript can run in the browser.
Release history
Dated milestones from Google's own pages, as of 15 November 2026.
| Date | Release | What changed |
|---|---|---|
| 9 Apr 2025 | ADK announced | Open source framework launched at Google Cloud Next 2025, with a CLI, web UI, evaluation and deployment to Vertex AI Agent Engine |
| 15 Oct 2025 | First community meeting | Listed in the adk-python README |
| 19 May 2026 | Python v2.0.0 GA | Workflow Runtime, graph, dynamic and collaborative workflows; breaking changes to agent API, events and session schema |
| 30 Jun 2026 | Go v2.0.0 GA | Module path moves to google.golang.org/adk/v2; workflow package; tool confirmation |
| 17 Aug 2026 | Python v2.7.1 | Earliest entry on the first page of GitHub releases |
| 21 Aug 2026 | TypeScript v2.0.0 GA | Graph workflows; SequentialAgent, ParallelAgent and LoopAgent deprecated but working |
| 27 Aug 2026 | Python v1.39.1 | Bug-fix backport to the 1.x line |
| 10 Sep 2026 | Python v2.9.0 | Followed by v2.9.1 (15 Sep) and v2.9.2 (18 Sep) |
| 25 Sep to 12 Nov 2026 | Python v2.10.0 to v2.14.1 | v2.10.0 (25 Sep), v2.11.0 (2 Oct), v2.12.0 (13 Oct), v2.13.0 (22 Oct), v2.14.0 (4 Nov), v2.14.1 (12 Nov) |
Staying on 1.x: the 2.0 page gives pinning commands for teams not ready to move: pip install "google-adk~=1.0", npm install @google/adk@^1.6.0 and go get google.golang.org/adk@v1. Google's pages do not publish an end-of-support date for the 1.x lines as of November 2026.
Limits and known constraints
- Session compatibility is one-directional: the Python README says sessions written by 2.0 can be read by ADK 1.28 and later but not by older 1.x versions. The 2.0 page says to update every reader before writing 2.0 sessions to a shared database.
- Silent behaviour changes in Python 2.0:
BaseAgentnow subclassesBaseNode, and overrides of_run_async_impl()orgenerate_content()are silently ignored. Move that logic toBeforeAgentCallbackandAfterAgentCallback. - Error handling affects retries: broad
except Exception:blocks disable automatic retries underRetryConfig, and catchingBaseExceptiontrapsNodeInterruptedError, which breaks human-in-the-loop pauses. - Go serialisation keys changed: Go versions before v2.2.0 wrote some new event fields under Go field names. Non-Go validators must accept both key forms.
- Feature maturity differs by language: the Java README marks "Evaluate Agents" as coming soon and carries a Pre-GA notice for a preview feature; Kotlin's ML Kit module is a
-betarelease; Kotlin's live conversation mode is experimental. - Development UI is not for production: ADK Web is for development and debugging only.
Security guidance
This page covers what ADK is, how it is versioned and where it runs. Hardening is covered in the Google Agent Development Kit Security guide, and the wider discipline in the Securing Agent Frameworks Guide. Google's own safety guidance asks developers to start with a risk assessment for the agent's capabilities, domain and deployment, to layer several controls rather than rely on one, to use plugins for policies shared across agents and callbacks for agent-specific checks, and to build tools that expose only the intended actions (for example a read-only flag or a table allowlist set in Tool Context). The configurable Gemini content filters for hate speech, harassment, sexually explicit and dangerous content are off by default, so turn on the ones your use case needs. Apply ADK updates promptly through your normal dependency process.
Practitioner checklist
- Record which ADK language and major version each service uses, and whether it is still on a pinned 1.x line.
- Before upgrading Python agents to 2.0, search for overrides of
_run_async_impl()andgenerate_content()and move them to callbacks. - Update every session reader (to Python 1.28 or later, and any custom schema) before any service writes 2.0 sessions to shared storage.
- Replace broad
except Exception:andBaseExceptionhandlers so retries and human-in-the-loop pauses keep working. - Choose Agent-Auth or User Auth per tool deliberately, and log user attribution when tools act as a service account.
- Turn on Tool Confirmation for tools that change data or spend money, and apply the Model Armor, Gemini as a Judge or PII Redaction plugin at runner level where they fit.
- Keep ADK Web and
.envAPI keys out of production; use a managed identity and a deployment target from the deployment page. - Estimate hosting cost from the Agent Platform pricing page (or Cloud Run or GKE pricing) before choosing a target.
Standards and references
- Google: Agent Development Kit documentation
- Google: ADK 2.0 overview and migration
- Google: ADK safety and security
- Google: ADK deployment options
- Google: ADK models
- Google: adk-python repository
- Google Developers Blog: Agent Development Kit announcement
- Google Cloud: Gemini Enterprise Agent Platform pricing
Related NHI Mgmt Group resources: Google Agent Development Kit Security, Gemini Enterprise Agent Platform, Google Cloud Agent Identity, Securing Agent Frameworks Guide.