Google ADK

Plugin-based harness integration for Google Agent Development Kit with budget enforcement and metrics tracking.

cascadeflow integrates with Google’s Agent Development Kit (ADK) through the BasePlugin system. Call enable() to get a plugin that plugs into Runner(plugins=[...]), keeping runtime measurement and enforcement close to the ADK execution flow instead of pushing it out to a separate proxy layer.

Install

pip install "cascadeflow[google-adk]"

Requires Python 3.10+.

Quick Start

import asyncio
from google.adk.agents import Agent
from google.adk.runners import Runner
from google.adk.sessions import InMemorySessionService
from google.genai.types import Content, Part

import cascadeflow
from cascadeflow.integrations.google_adk import GoogleADKHarnessConfig, enable

cascadeflow.init(mode="observe")

# Enable harness plugin
config = GoogleADKHarnessConfig(
    fail_open=True,
    enable_budget_gate=True,
)
plugin = enable(config=config)

# Create ADK agent
agent = Agent(
    name="research_agent",
    model="gemini-2.5-flash",
    instruction="You are a helpful research assistant.",
)

# Run with plugin
session_service = InMemorySessionService()
runner = Runner(agent=agent, plugins=[plugin])

async def main():
    with cascadeflow.run(budget=0.50) as session:
        user_content = Content(parts=[Part(text="Explain cascadeflow")])
        async for event in runner.run_async(
            session_id="test",
            user_id="user-1",
            new_message=user_content,
        ):
            pass  # Process streaming events

        print(session.summary())

asyncio.run(main())

Configuration

config = GoogleADKHarnessConfig(
    fail_open=True,          # Continue on harness errors
    enable_budget_gate=True, # Enforce budget caps
)

Supported Gemini Models

Model	Input $/1M	Output $/1M	Energy Coeff
gemini-2.5-flash	$0.15	$0.60	0.30
gemini-2.5-pro	$1.25	$10.00	1.20
gemini-2.0-flash	$0.10	$0.40	0.25
gemini-1.5-flash	$0.075	$0.30	0.20
gemini-1.5-pro	$1.25	$5.00	1.00

Budget Enforcement

When budget is exceeded in enforce mode, the plugin returns an LlmResponse with error_code="BUDGET_EXCEEDED". The ADK runner handles this as a graceful stop.

Why This Integration Matters

• ADK runners can stay framework-native while gaining runtime governance
• Budget control and traces apply at the actual execution boundary
• The integration keeps the in-process latency advantage intact

Limitations

• Tool gating is not applied (intentional design choice — ADK manages tool execution internally)
• Model switching depends on ADK’s model configuration

Example on GitHub: integrations/google_adk_harness.py