Skip to content
GkPy

Google ADK (Python)

Google Agent Development Kit for Python: Gemini-first agents with Workflow orchestration, MCP, A2A, the adk CLI and deployment to Agent Engine, Cloud Run and GKE.

Google’s Agent Development Kit for Python. Build agents on Gemini (or other models through LiteLlm, AnthropicLlm and the OpenAI classes), compose them with Workflow graphs, connect them to MCP servers and other agents over A2A, and deploy them with the adk CLI.

Install

Terminal window
pip install google-adk

Version

2.13.0 · Python 3.11+

Best for

Gemini-first agents, Google Cloud workloads, Vertex AI Agent Engine.

Terminal window
pip install google-adk
adk create my_agent --model gemini-3.8-flash # asks for an API key or a Google Cloud project
adk run my_agent # chat in the terminal
adk web . # or in the Dev UI at http://127.0.0.1:8000

The same agent from a script:

import asyncio
from google.adk.agents import LlmAgent
from google.adk.runners import InMemoryRunner
agent = LlmAgent(
name="assistant",
model="gemini-3.8-flash",
instruction="You are a helpful assistant.",
)
async def main():
runner = InMemoryRunner(agent=agent, app_name="demo")
await runner.session_service.create_session(app_name="demo", user_id="u1", session_id="s1")
events = await runner.run_debug("What is 15 + 27?", user_id="u1", session_id="s1")
print(events[-1].content.parts[0].text)
asyncio.run(main())

Package releases of google-adk 2.x, with the changes that affect code in this guide. Full notes are in the ADK changelog.

VersionReleasedSummary
2.13.02026-10-22Telemetry defaults to the OpenTelemetry GenAI schema: the root span is invoke_workflow (was invocation) and there is no call_llm span; set ADK_TELEMETRY_SCHEMA_VERSION_OPT_IN=otel_semconv_1_36 to keep the old one. OpenAPIToolset and RestApiTool refuse private, loopback and link-local targets unless allow_private_network_access=True. adk eval exits non-zero when a case fails. Twelve experimental features become stable, among them AuthenticatedFunctionTool and ContextCacheConfig, and ADK_SUPPRESS_EXPERIMENTAL_FEATURE_WARNINGS now silences every feature warning. New, experimental: MongoDbSessionService and MongoDbMemoryService, ClassifierNode with LlmClassifier, BaseMemoryService.delete_memory (implemented by InMemoryMemoryService), ForceEscalationPlugin, Gemini speech-to-text and text-to-speech with barge-in for CascadeLive. Up to 8x faster turns for agents given a model name string.
2.12.02026-10-13Python 3.10 is no longer supported (requires-python >=3.11). after_tool_callback can screen tool output. require_confirmation on @node and FunctionNode. adk deploy agent_engine --service_account. ADK moves to google-cloud-aiplatform 2.2 (the agentplatform SDK). Built-in search tools work alongside sub-agents on Gemini 3. Removed: the workflow HITL helpers has_request_input_function_call and get_request_input_interrupt_ids.
2.11.02026-10-01abort_signal on Runner, Workflow and nodes; /run_sse cancels the run when the client disconnects. Tool nodes in workflows pause for confirmation like LlmAgent tools. ModelConsultTool. SQLite memory service (sqlite://). Opt-in MCP SDK 2.x path.
2.10.02026-09-24OpenAI models move to google.adk.integrations.openai with a google-adk[openai] extra (google.adk.labs.openai re-exports them). Experimental skill lifecycles, MongoDB toolset, efficiency metrics for evals. Behaviour changes: instruction templating leaves ${var} and \{var} alone; AgentEvaluator.evaluate raises ValueError when no case runs.
2.9.0 – 2.9.22026-09-10 to 09-18FallbackModel for automatic failover between models. LiveKit runner. Workflow graphs from YAML. MCP SDK 2.x support. Behaviour changes: a failed workflow node re-runs on resume; InMemorySessionService raises SessionNotFoundError when appending to an unknown session.
2.8.02026-08-25Model Armor guardrail plugin. Native task mode for RemoteA2aAgent. ADK_MAX_LLM_CALLS environment variable.
2.7.0 – 2.7.12026-08-13 to 08-17Models declare their own capabilities (output schema with tools where supported). Tools can return media. Jinja2 instruction templates via inject_session_state(use_jinja2=True). Breaking: pyarrow moved from the gcp extra to the new bigquery-analytics extra.
2.6.0 – 2.6.32026-07-29 to 08-07Breaking: file artifacts are namespaced by app; LangGraph integration requires the patched async runtime. adk deploy agent_engine --extra_packages.
2.5.02026-07-16to_mcp_server serves an ADK agent over MCP. Cloud Run sandbox code executor. Strict input-schema validation for LlmAgent workflow nodes.
2.4.02026-07-07ManagedAgent (Managed Agents API). A Workflow can be used as a tool. Session TTL in VertexAiSessionService. OpenAI Responses model in labs.
2.3.02026-06-17GEPARootAgentOptimizer. GCS toolset. ContextCacheConfig.create_http_options. Core and CLI move to the GOOGLE_GENAI_USE_ENTERPRISE parameters.
2.2.02026-06-04AutoTracingPlugin, native gen_ai.client.* OpenTelemetry metrics, --trigger_sources for Agent Engine deployments, RubricBasedMultiTurnTrajectoryEvaluator.
2.1.02026-05-23Sandboxes from templates and snapshots; telemetry and transcription fixes.
2.0.02026-05-19General availability of ADK 2: the graph workflow engine (Workflow, nodes, edges), with SequentialAgent / ParallelAgent / LoopAgent deprecated in its favour.

Earlier 1.x releases (November 2025 to May 2026) are covered in the upstream changelog.

Other languages: Go · TypeScript.