Sandbox

• 

  sanbox preview

• 

  Agent Architecture Diagram with Custom Agents

Inspiration

Campaign planning is still painfully fragmented.
Creative teams juggle slides, spreadsheets, and group chats, while AI tools stay linear and chat based.

We wanted to build something that thinks in space, a system that understands the relationship between ideas, budgets, and execution, and can visualize it all as a living map.

That’s how Sandbox was born, a space where you can describe an idea in plain language, and the system instantly generates how to bring it to life.

What It Does

Sandbox turns campaign ideas into orchestrated, dynamic systems.
It uses spatial intelligence to represent campaigns as a live node based canvas, where every node represents an actionable element like content, distribution, budget, or outcomes, all connected by intelligent logic.

You start with a simple prompt, for example:

“Launch a lifestyle campaign for a sustainable apparel brand.”

From there, Sandbox:

• Parses intent using an AI agent.
  
• Generates target audiences, creative tone, and budget split.
  
• Builds an interactive orchestration map.
  
• Adapts dynamically as you continue the conversation.

Every campaign in Sandbox feels alive, evolving as you think, plan, and iterate.

How We Built It

Sandbox is powered by Google Cloud Run, Google ADK (Agent Development Kit), and the AG UI (Agentic UI) protocol.

We designed a multi agent architecture where:

• The Campaign Generator Agent creates structured campaign logic from natural language inputs.
  
• The Deep Research Agent gathers contextual data (audience, trends, cost benchmarks) using callbacks, Google Search, and programmatic loops, without falling into uncontrolled infinite loops.
  
• The frontend is built with React and a node based canvas that visualizes agent outputs in real time. AG UI allows both the UI and the agent to operate on a shared state, so the interface updates automatically as the system learns or adapts.

Key Cloud Run components:

• ADK powered microservices handle AI reasoning and orchestration.
  
• Custom callback services ensure structured data output.
  
• Cloud Run containers scale agents independently for performance and cost efficiency.

Working with AG UI

There is almost no documentation or examples, so building from scratch required extensive trial and error to sync the frontend and agent states.

Structured agent output

LLMs often hallucinate or drift from schema.
We solved this using custom callbacks and rule based parsers to enforce structured responses.

Agent loops

Our Deep Research Agent needed iterative search capabilities, but pure LLM loops risked going infinite.
We combined AI reasoning with controlled program logic to manage loop depth dynamically.

Latency on orchestration updates

Balancing responsiveness and scalability in Cloud Run required tuning cold start parameters and microservice timeouts.

What We Learned

• How to combine AI reasoning with deterministic logic for reliable orchestration.
  
• How AG UI enables generative user interfaces (where the frontend itself becomes part of the reasoning loop).
  
• That spatial intelligence (representing ideas as nodes, not lists) unlocks a much more intuitive way to plan creative systems.
  
• How Cloud Run makes multi agent orchestration modular, serverless, and surprisingly affordable.

What’s Next

We are expanding Sandbox into a full creative operating system, one that helps founders, strategists, and marketers ideate, plan, and execute campaigns collaboratively.

Future plans include:

• Multi user live collaboration on the orchestration canvas.
  
• Auto learning templates for different campaign types.
  
• Deeper integration with Google Ads, Analytics, and YouTube APIs.

Our long term vision is to make Sandbox the spatial brain for creative strategy, where AI does not just answer, it orchestrates.

Team

Creative Technologist: Wafi Abd
AI Engineer: Abhishek Kumar

Built for the Google Cloud Run Hackathon 2025.

Built With

• ag-ui
• cloudrun
• nextjs
• protocol
• react