Built SmartDesk, a multi-agent AI system using Google ADK to automate tasks, scheduling, notes, and research end-to-end.

When I signed up for the Google Cloud Gen AI Academy APAC 2026 Hackathon, the problem statement was clear:

"Build a multi-agent AI system that helps users manage tasks, schedules, and information by interacting with multiple tools and data sources."

Simple enough on paper. But building it taught me more about real-world AI systems than anything I'd done before. This is the full story — what I built, how I built it, the mistakes I made, and what I learned.

The Problem I Chose to Solve

Here's a stat that hit me hard when I was planning this project:

The average knowledge worker spends 2.5 hours every day just organizing their work — not doing it. Switching between Notion for tasks, Google Calendar for scheduling, a notes app for summaries, and a browser tab for research.

Every tool is smart on its own. But none of them talk to each other intelligently. You're the human glue between them.

I wanted to build something that eliminates that. One conversational interface that handles everything end-to-end. You describe what you need in plain English — and a team of AI agents handles the rest.

That's SmartDesk — your personal AI Chief of Staff.

The Idea: One Prompt, Four Agents

Instead of building a single AI assistant, I designed SmartDesk as a team of specialist agents, each owning one responsibility:

✅ Task Agent — creates and manages tasks with priorities and deadlines
📅 Calendar Agent — schedules events and blocks focus time with conflict detection
📝 Notes Agent — captures summaries and action items automatically
🔍 Research Agent — answers knowledge questions via Wikipedia

A single prompt like this:

"I have a product launch deadline this Friday. Block 2 hours of focus time daily, create tasks for design, development, and testing milestones, and save a project summary note."

…triggers all four agents to execute their jobs in sequence, passing context to each other, and returns one clean consolidated response.

No app switching. No manual entry. Just results.

How It Works — The Architecture

SmartDesk is built on Google ADK (Agent Development Kit) using a SequentialAgent pattern.

Here's the full flow:

User Input (natural language)
↓
Root Agent (smartdesk_root)
→ saves prompt to shared state
↓
Sequential Workflow (smartdesk_workflow)
→ Task Agent writes task_output to state
→ Calendar Agent reads task_output, writes calendar_output
→ Notes Agent reads all outputs, writes notes_output
→ Research Agent checks if knowledge lookup needed
→ Response Formatter synthesizes everything
↓
Final Response to User

The key insight is shared state. Every agent reads from and writes to tool_context.state — a shared memory dictionary that flows through the entire pipeline. This is what makes true multi-agent coordination possible. Each agent builds on the work of the one before it.

The Database

All data is persisted in SQLite across four tables:

tasks → title, priority, deadline, status

events → title, start_time, end_time

notes → content, tags, created_at

agent_logs → agent_name, action, timestamp ← my favorite

That last table — agent_logs — logs every single agent action with a timestamp. During demos, watching it fill up in real-time is the most compelling way to show judges that genuine multi-agent coordination is happening.

The Tech Stack
Layer Technology Why I chose it
Agent Framework Google ADK 1.14.0 Native Google Cloud integration, built-in Cloud Run deployment
LLM Gemini 2.5 Flash (Vertex AI) Fast inference for multi-step pipelines, no API key needed with ADK
Database SQLite Zero configuration, perfect for hackathon demos
External Knowledge LangChain + Wikipedia Bridges ADK with the LangChain tool ecosystem
Deployment Google Cloud Run Serverless, one-command deploy via adk deploy cloud_run
Auth Google IAM Service Account Least-privilege security
Building It — Step by Step
Setting Up the Environment

I used Google Cloud Shell — a browser-based development environment with everything pre-installed. No local setup required.

cd && mkdir smartdesk && cd smartdesk
uv venv && source .venv/bin/activate
uv pip install -r requirements.txt

The requirements.txt was simple:

google-adk==1.14.0
langchain-community==0.3.27
wikipedia

One thing I learned the hard way: wikipedia==1.4.8 doesn't exist. The version pin breaks the install. Remove the version and let pip grab the latest.

Writing the Agents

Each agent in Google ADK is defined with a name, model, instruction, tools, and an output_key:

task_agent = Agent(
name="task_agent",
model=model_name,
description="Creates and manages tasks in the database.",
instruction="""
You are the Task Manager Agent for SmartDesk.
Based on the user's PROMPT, break it down into clear tasks.
Create each task using the create_task tool.
PROMPT: { PROMPT }
""",
tools=[create_task, get_all_tasks],
output_key="task_output"
)

The { PROMPT } syntax injects shared state values into agent instructions.

The SequentialAgent
smartdesk_workflow = SequentialAgent(
name="smartdesk_workflow",
sub_agents=[
task_agent,
calendar_agent,
notes_agent,
research_agent,
response_formatter
]
)

The SequentialAgent handles execution order automatically.

Key Challenges I Faced

1. SQLite Path on Cloud Run

Error:

"Fail to load 'smartdesk' module. unable to open database file"

Fix:

Before

DB_PATH = "smartdesk.db"

After

DB_PATH = "/tmp/smartdesk.db"

1. Service Account Name Too Short

Error:

ERROR: The account ID "lab2" does not have a length between 6 and 30.

Fix: renamed to smartdesk-agent.

1. Git Push Rejection

Error:

error: failed to push some refs — Updates were rejected because the remote contains work that you do not have locally.

Fix:

git pull --rebase origin main
git push
The Demo — What Judges See

Prompt:

"I have a product launch deadline this Friday. Block 2 hours daily, create tasks for design, dev, and testing, and save a project summary note."

Execution trace:

smartdesk_root → task_agent → calendar_agent → notes_agent → research_agent → response_formatter

Then showing agent_logs filling in real time — proof of real multi-agent coordination.

What I Learned
Shared state is the backbone of multi-agent systems.
Observability matters more than output.
Google ADK makes deployment trivially easy.
Start with the demo, work backward.
SQLite is underrated for AI demos.
What's Next for SmartDesk
🔌 Real MCP integrations — Google Calendar API, Notion, Slack
🧠 Long-term memory — ChromaDB
🗣️ Voice interface — faster-whisper
📊 Analytics dashboard — agent_logs insights
🤝 Team mode — multi-user collaboration
Try It Yourself
GitHub: https://github.com/arpitpandey0307/smart-desk
Live Demo: https://smartdesk-agent-456703325610.us-central1.run.app
Demo Video: https://drive.google.com/file/d/17i04sEihCSWR1J7ioeOPM0dH4MoaxzQy/view
Final Thoughts

Building SmartDesk taught me that the real power of multi-agent AI isn't in any single agent — it's in the coordination layer between them.

The moment you design agents that pass context to each other, you stop building tools and start building systems.

Built with Google ADK, Gemini 2.5 Flash, Vertex AI, SQLite, and Google Cloud Run.