Google Adk App

Google ADK — App Pattern

When to Use App vs Plain Agent

Import

from google.adk.agents import Agent
from google.adk.apps import App, ResumabilityConfig

Basic App

from google.adk.agents import Agent
from google.adk.apps import App

root_agent = Agent(
    name="my_agent",
    model="gemini-2.5-flash",
    instruction="You are a helpful assistant.",
    tools=[],
)

app = App(
    name="my_app",
    root_agent=root_agent,
    plugins=[],
)

App with Plugins

from google.adk.plugins import ContextFilterPlugin

app = App(
    name="my_app",
    root_agent=root_agent,
    plugins=[
        ContextFilterPlugin(num_invocations_to_keep=3),
    ],
)

Available Plugins

ContextFilterPlugin

Keeps only the last N invocations in context (saves tokens):

from google.adk.plugins import ContextFilterPlugin

plugin = ContextFilterPlugin(
    num_invocations_to_keep=5,  # Only last 5 turns sent to model
)

Event Compaction

Summarizes old events to manage session size over long conversations:

from google.adk.apps import App
from google.adk.apps.compaction import EventsCompactionConfig
from google.adk.apps.llm_event_summarizer import LlmEventSummarizer

app = App(
    name="my_app",
    root_agent=root_agent,
    events_compaction_config=EventsCompactionConfig(
        summarizer=LlmEventSummarizer(model="gemini-2.5-flash"),
        compaction_interval=10,  # Compact after every 10 invocations
        overlap_size=2,  # Keep 2 invocations of overlap
    ),
)

How Compaction Works

1. After compaction_interval new invocations accumulate
2. The summarizer generates a summary of older events
3. Old events are replaced with a CompactedEvent containing the summary
4. overlap_size ensures continuity between compacted and live events

Resumability

For long-running operations that may timeout or fail:

from google.adk.apps import App, ResumabilityConfig

app = App(
    name="my_app",
    root_agent=root_agent,
    resumability_config=ResumabilityConfig(
        is_resumable=True,
    ),
)

Resumability Requirements

• Tool calls must be idempotent (safe to retry)
• Only guarantees at-least-once execution on resume
• Temporary/in-memory state is lost on resumption

Context Cache Config

Cache static instruction content for cost savings:

from google.adk.agents import Agent
from google.adk.agents.context_cache_config import ContextCacheConfig

agent = Agent(
    name="cached_agent",
    model="gemini-2.5-flash",
    instruction="Very long system instruction...",
    context_cache_config=ContextCacheConfig(
        enabled=True,
        min_tokens_to_cache=2048,
    ),
)

App Discovery by CLI

The CLI finds apps the same way it finds agents:

# agent.py — CLI auto-discovers `app` variable
from google.adk.agents import Agent
from google.adk.apps import App

root_agent = Agent(name="my_agent", model="gemini-2.5-flash", instruction="...")

app = App(
    name="my_app",
    root_agent=root_agent,
    plugins=[...],
)

When both root_agent and app exist, the CLI uses app.

App Parameters

Key Rules

• App name must be a valid Python identifier and cannot be "user"
• When app variable exists in agent.py, CLI uses it over root_agent
• Plugins run on every invocation — keep them lightweight
• Compaction trades detail for token efficiency — tune compaction_interval
• Resumability requires idempotent tools

Related Skills

• google-adk-session — Session services (used by App)
• google-adk-memory — Memory services (used by App)
• google-adk-deploy — Deploying apps to production