DEV Community

Cover image for Zero to Production AI Agents with Gemini 3 + ADK: Days 1-4 + Day 8
Cless
Cless

Posted on Edited on

Zero to Production AI Agents with Gemini 3 + ADK: Days 1-4 + Day 8

#ai #agents #adventofagents #gemini

Advent of Agents - Days 1-4 + Day 8: Zero to Production AI Agents

Source: Google Cloud Advent of Agents Season 1 – December 2025 (Winter Holiday Edition)
25 days. Zero to Production-Ready AI Agents on Google Cloud. Every day < 5 min to try.

Table of Contents

Day 1: Launch Initiative

Original: adventofagents.com/2025/12/01 | Summary: 25 days. Zero to Production-Ready AI Agents on Google Cloud.

Advent of Agents 2025 starts NOW!

Every day until Christmas:

  • ⚡ One feature, < 5 min to try
  • 💻 Copy-paste commands that actually work
  • 📖 Links to docs

From quickstart agents to multi-agent orchestration. From local dev to production deployment.

What we are covering:

📚 Recommended Reading: Introduction to Agents whitepaper – released as part of the 5 Days of Agents with Kaggle.

More info here

Day 2: Hello World with YAML

Original: adventofagents.com/2025/12/02 | Summary: Build your first AI agent with Gemini 3 in under 5 minutes without writing a single line of code.

  • 4 lines of text = 1 working AI agent.
  • No coding. Just YAML.
  • A working AI agent powered by Gemini 3
  • Google Search integration

What you get with 4 lines of YAML:

  • A working AI agent powered by Gemini 3
  • Google Search integration
  • Ready to deploy
  • Zero programming knowledge needed

Once you have your YAML agent, you can add:

  • Built-in tools (google_search, code_execution)
  • Custom Python tools (when you're ready)
  • Sub-agents for complex workflows
  • Multi-agent orchestration

Code & Commands 

uvx --from google-adk adk create --type=config my_agent
uvx --from google-adk adk web my_agent/
Enter fullscreen mode Exit fullscreen mode

Resources

Day 3: Gemini 3 + ADK

Original: adventofagents.com/2025/12/03 | Summary: Build a powerful AI Agent using Gemini 3 and ADK with native support for Google Search grounding, computer use, and real-time streaming.

Gemini 3 is Google's most intelligent model yet. In just a few minutes, you can build a powerful AI Agent using Gemini 3 and ADK with native, day-one support for its most advanced capabilities.

What you get with ADK and Gemini 3:

  • Google Search grounding: Native access to real-time web data
  • Computer use: Agents that navigate and interact with UIs
  • Live API: Real-time streaming for voice and video agents
  • Native observability: Full visibility into Gemini calls, tool use, and agent reasoning

1. Setup the project

One liner with Agent Starter Pack:

uvx agent-starter-pack create -y --api-key <YOUR_GEMINI_API_KEY> my_agent
Enter fullscreen mode Exit fullscreen mode

or Using ADK CLI:

uv init
uv add google-adk
uv add google-genai
export GOOGLE_API_KEY="YOUR_API_KEY"
source .venv/bin/activate
adk create my_agent
Enter fullscreen mode Exit fullscreen mode

2. Download sample and run locally 

curl 'https://raw.githubusercontent.com/GoogleCloudPlatform/devrel-demos/refs/heads/main/ai-ml/agent-labs/gemini-3-pro-agent-demo/my_agent/agent.py' > my_agent/agent.py
adk web
Enter fullscreen mode Exit fullscreen mode

Minimal agent.py from the demo:

from google.adk.agents import Agent
from google.adk.tools import google_search

root_agent = Agent(
    name="search_agent",
    model="gemini-3-pro-preview",
    description="A helpful assistant that can search Google.",
    instruction="""You are a helpful assistant with access to Google Search.
    If the user asks a question that requires current information or facts, use the 'google_search' tool.
    Always cite your sources implicitly by providing the answer clearly based on the search results.
    """,
    tools=[google_search]
)
Enter fullscreen mode Exit fullscreen mode

Resources

Day 4: Source-Based Deployment

Original: adventofagents.com/2025/12/04 | Summary: Deploy your agents directly from source code with Agent Engine - no more serialization headaches.

Agent Engine now supports source-based deployment. Your source code deploys directly to production, eliminating common deployment friction.

No more dealing with:

  • Pickle errors
  • Serialization nightmares
  • Module path headaches

What runs locally now runs in production.

Get started with Agent Starter Pack:

For a new project:

uvx agent-starter-pack create my-agent -a adk_base -d agent_engine
cd my-agent && make deploy
Enter fullscreen mode Exit fullscreen mode

Already have an ADK agent? Use enhance:

uvx agent-starter-pack enhance --d agent_engine
Enter fullscreen mode Exit fullscreen mode

Then trigger the deployment with:

make deploy
Enter fullscreen mode Exit fullscreen mode

Resources

Day 8: Effective Context Management with ADK Layers

Original: adventofagents.com/2025/12/08 | Summary: ADK Context design thesis: context as a compiled view

ADK Context Engineering Infographic

Stop stuffing spaghetti context into your agent's LLM.

The "append-everything" strategy is a one-way ticket to latency spirals and "lost in the middle" hallucinations. Google ADK shifts the paradigm by treating context as a compiled view (not a giant string).

Instead of shoveling raw history into the window, ADK uses a pipeline of processors to dynamically filter, compact, and format a clean "working context" derived from a structured, durable session state.

Real engineering requires granular control, not just a bigger context window.

ADK's tiered architecture separates storage from presentation, allowing you to externalize massive files as Artifacts (using a handle pattern) and retrieve long-term data via searchable Memory only when strictly necessary.

ADK operationalizes Context Caching by enforcing a clear separation between "Static Instructions" (invariant policies and schemas) and "Turn Instructions" (dynamic, controller-owned steering).

As described in Day 3 of our Kaggle 5 Day intensive course, Agents are largely "context management". This is how you do it.

Code: Static Context Policy + Context Caching 

# An Example of Static Context Policy 
from google.adk.apps import App
from google.adk.agents import Agent
from google.adk.agents.context_cache_config import ContextCacheConfig

STATIC_POLICY_HEADER = """You are a strict policy assistant for internal compliance Q&A.

Follow this exact JSON schema in every response:
{"answer": str, "citations": [str], "confidence": float}

Safety:
- Never provide medical or legal advice; refuse with a brief explanation.
- Never invent policy numbers or sections; ask for the missing reference.

Style:
- Use short sentences.
- Prefer active voice.
- If uncertain, say so and request the missing input.

Tools:
- search: use for public web facts.
- bq: use for internal policy tables (read-only).
"""

agent = Agent(
    name="policy_agent",
    static_instruction=STATIC_POLICY_HEADER,
    instruction="Default: be concise and include at most two citations."
)

app = App(
    name="policy_qa_app",
    context_cache_config=ContextCacheConfig(
        ttl_seconds=3600,     # cache the header for 1 hour
        cache_intervals=5,    # force a refresh every 5 requests (guardrail)
        min_tokens=1000       # only cache if header is “worth it”
    ),
    root_agent=agent
)
Enter fullscreen mode Exit fullscreen mode

Code: Runtime controller generating turn instructions 

from dataclasses import dataclass
from typing import Optional, Tuple

@dataclass
class SteeringInputs:
    goal: str
    style: str = "concise"
    max_cites: int = 2
    tenant_hint: Optional[str] = None
    corrective: Optional[str] = None
    confidence_range: Tuple[float, float] = (0.6, 0.9)

def build_turn_instruction(s: SteeringInputs) -> str:
    parts = [
        f"Goal: {s.goal}",
        f"Style: {s.style}",
        (
            "Constraints: "
            f"include at most {s.max_cites} citations; "
            "refuse medical/legal advice; "
            "if info is missing, ask one targeted question; "
            f"return 'confidence' between {s.confidence_range[0]} and {s.confidence_range[1]}."
        )
    ]
    if s.tenant_hint:
        parts.append(f"Tenant: {s.tenant_hint}")
    if s.corrective:
        parts.append(f"Correction: {s.corrective}")
    return "\n".join(parts)
Enter fullscreen mode Exit fullscreen mode 
def chat(session_id: str, user_message: str, ui_style: str | None = None):
    intent = route_intent(user_message)
    goal = INTENT_TO_GOAL.get(intent, f"Answer the user: {user_message[:120]}")

    style = ui_style or get_flag(session_id, "style", default="concise")
    max_cites = get_flag(session_id, "max_citations", default=2)
    tenant_hint = get_tenant_hint(session_id)
    corrective = get_last_validation_error(session_id)

    turn_instruction = build_turn_instruction(
        SteeringInputs(
            goal=goal,
            style=style,
            max_cites=max_cites,
            tenant_hint=tenant_hint,
            corrective=(f"Your last reply failed validation: {corrective}. Fix it this turn." if corrective else None)
        )
    )

    agent.instruction = turn_instruction
    response = agent.run(user_message=user_message)
    validate_and_record(session_id, response)
    return response
Enter fullscreen mode Exit fullscreen mode

Resources

What's Next?

This combines Days 1-4 + 8. Full Season 1 is 25 days: observability, MCP, Live API, A2A, A2UI, Gemini Enterprise, security, durable execution with Restate, and finale.

Follow for the rest, and bookmark adventofagents.com for the permanent archive.

Top comments (0)