Google ADK (Python)

ADK 2.0 moved from a hierarchical agent executor to a graph-based execution engine. Agents, tools and plain functions are nodes in a Workflow, and edges define routing, parallel branches, loops and retries. An agent is itself a node, so graphs can mix deterministic code steps with LLM reasoning and can be nested.

For less structured work, a coordinator agent can delegate to subagents, each set to a collaboration mode: chat (full user interaction), task (asks for clarification, then returns to the parent) or single-turn (no user interaction, can run in parallel). Older template agents (sequential, parallel, loop) remain available.

Conversation context is split into a Session (the event history of one conversation), State (data scoped to that session) and a Memory service that can be searched across sessions, with in-memory, database and Vertex AI backed implementations. Resumable apps can pick up an interrupted workflow where it stopped.

Three documented routes. Tool confirmation (experimental) wraps a tool with require_confirmation=True for a yes/no answer, or requests a structured response with a prompt; the answer can come from the ADK web UI dialog or be posted remotely through the ADK server's REST API. Known limitation: it does not work with DatabaseSessionService or VertexAiSessionService. In ADK 2.0 graph workflows, a node can yield RequestInput to pause until a person replies, and the reply becomes the next node's input. Subagents in task mode can also stop to ask the user for clarification before returning to their coordinator.

pip install google-adk

# my_agent/agent.py (project from `adk create my_agent`); run: adk run my_agent
from google.adk import Agent, Workflow
from google.adk.tools import FunctionTool

def refund(order_id: str) -> dict:
    """Issue a refund for an order."""
    return {"status": "refunded", "order_id": order_id}

triage = Agent(name="triage", model="gemini-2.5-flash",
               instruction="Summarize the customer's problem in one line.")
resolver = Agent(name="resolver", model="gemini-2.5-flash",
                 instruction="Resolve the issue; call refund only when needed.",
                 tools=[FunctionTool(refund, require_confirmation=True)])

root_agent = Workflow(
    name="root_agent",
    edges=[("START", triage, resolver)],
)

• Python 3.10 or later is required; the README recommends installing with its constraints file for your Python version.
• Gemini models need GOOGLE_API_KEY (or Vertex AI settings) in the project's .env.
• ADK 2.0 changed the event schema (node_info, output) and made agents graph nodes; custom session stores and overrides of _run_async_impl() need migration, and old overrides are silently ignored.
• Broad except Exception blocks in tools disable 2.0 automatic retries, and catching BaseException breaks HITL pauses.
• Tool confirmation is experimental and unsupported with DatabaseSessionService and VertexAiSessionService.
• To stay on 1.x, install with the compatible-release pin pip install "google-adk~=1.0".

Yes. McpToolset lets agents use MCP server tools, and to_mcp_server() turns an agent into an MCP server that clients such as Claude Code can call.

Partly. Python has to_a2a() for exposing agents and RemoteA2aAgent for calling remote ones, but the docs mark A2A support as experimental.

The framework is Apache-2.0. Deploying to Google Cloud's managed Agent Runtime is a paid service once usage exceeds its no-cost tier.

No. The README says ADK is optimized for Gemini but model-agnostic and deployment-agnostic.

Python 2.0 (GA on 2026-05-19) added a graph-based Workflow runtime and collaborative agent modes, and made agents nodes in the graph, with some breaking changes for 1.x code.