This post is my submission for DEV Education Track: Build Multi-Agent Systems with ADK.
"A lot of AI demos look impressive but still hide the actual orchestration. I wanted to make those handoffs deeply explicit."
This post is my official submission for the DEV Education Track on Building Multi Agent Systems.
What I Built
I built Multi Agent Studio, a highly visual web application that turns one broad request into a fully explorable multi agent workflow.
Instead of asking a single hidden model to do everything in one pass, the app intelligently splits the job across four focused roles.
First, Atlas Story handles the heavy planning and framing.
Second, Signal Curator is strictly responsible for evidence gathering and contradiction mapping.
Third, Vector Ops manages the dense execution packaging.
Fourth, Relay Console is thoughtfully used for human review, escalation, and operator readiness.
The application is specifically designed for users who do not want to babysit a giant prompt.
They can dynamically choose a workflow template, edit their core objective, and deeply inspect the full chain of handoffs, artifacts, and recommendations.
Live Demo and Source Code
You can view the full source code directly embedded below.
aniruddhaadak80
/
ai-agents-duel
An interactive, narrative-driven AI Agent command center featuring a whimsical Digital Sketchbook UI, live neural feeds, and agent dueling.
Explore the Application
A fully functional Next.js multi agent system that elegantly turns one broad user request into a visible planner, researcher, builder, and reviewer workflow.
Live app: https://ai-agents-duel.vercel.app
GitHub: https://github.com/aniruddhaadak80/ai-agents-duel
This project was carefully rebuilt around the core ideas from the DEV Education track on building multi agent systems. The primary focus is squarely on deep specialization, structured orchestration, explicit handoffs, and rigid review gates. Instead of one giant prompt, the app gives users an entire workflow library, precise agent controls, a visual review queue, and fully inspectable run details.
Extensive Topics and Tags
Here is a look at the major themes covering this repository AI Agents, Build Multi Agents, Gemini 3 Flash, Nextjs, Vercel, Generative AI, Automation, LLM Orchestration, TypeScript Engineering, React 19, Human In The Loop, Agentic Workflows, Google AI Studio, Developer Tools.
Visual Storyboard and Interface
Here is a look at the live dynamic application.
You…
You can also visit the live application on Vercel to try it yourself right now.
Visual Walkthrough
You can browse the natively available pre built templates directly in the Workflow Library.
You can deeply inspect the exact output and internal thought process in the Run Detail view.
The Mission Board provides a genuinely clear, categorized view of your active tasks and agent assignments in real time.
How the system works
The Core Problem
Users typically type one request, get one answer, and simply cannot tell which role should have handled what.
They do not see where the system confidence silently dropped.
They also do not know exactly when human review is deeply needed or what can be acted on immediately.
By making handoffs explicit, we turn abstract AI operations into a tangible, observable assembly line.
How It Works Under The Hood
The application relies on a workflow first operational model.
Users pick from rich templates like Launch Campaign Studio, Ops War Room, Founder Decision Desk, and Ship Feature Relay.
Each continuous run produces transparent stage ownership across the designated agent team.
It carefully provides a concrete review state, specific agent contributions, and predictive next step recommendations.
Finally, an operator brief is actively generated for incredibly fast human action.
The core orchestration engine currently runs entirely in memory for effortless local setup and seamless Vercel deployment.
If a GEMINI_API_KEY is configured, the selected agent intelligently enriches the final run output utilizing the brand new Gemini 3 Flash Preview model.
Architecture Pattern
The entire system follows a rigorous multi agent pattern logically broken into sequential steps.
First, the Planner accurately scopes the core mission.
Next, the Researcher rigorously gathers context and maps out contradictions.
Then, the Builder systematically converts the raw work into an execution package.
Finally, the Reviewer thoughtfully checks confidence and firmly gates publication.
This architecture keeps the user experience wonderfully simple while still thoroughly proving the massive benefits of true specialization.
Technology Stack
The resilient foundation heavily relies on Next.js 16 and React 19 for a fast modern interface.
Everything is securely and robustly typed with TypeScript.
The styling remains exceptionally lightweight by only using Pure CSS.
The robust Google GenAI SDK reliably manages the artificial intelligence communication layer.
The whole rapid ecosystem is smoothly hosted and delivered on Vercel.
What Makes It Truly Different
I focused heavily on deep usability and structure instead of just raw aesthetics.
The app inherently includes workflow presets for common professional jobs and a real time mission board.
It elegantly features review queue actions alongside granular operator controls for autonomy tuning.
There are also beautifully built in agent pause and resume commands.
It provides an optional Gemini upgrade path without ever breaking standard local runs.
One main architectural challenge was actively avoiding a fake looking multi agent wrapper.
Many early demos simply stop at a stylish dashboard without technical depth.
I strongly wanted the internal model to be robust enough that the visual UI felt deeply justified.
That logically meant entirely rebuilding the internal run store around rigid workflows, stage ownership, structured artifacts, and clear review states instead of just random summaries.
Another prominent engineering challenge was cleanly keeping the project universally easy to deploy.
I effectively chose an in memory backend so the complete experience works immediately on Vercel.
This decision beautifully preserves a wonderfully clean path to add database persistence later.
Key System Learnings
Multi agent user experiences firmly become exponentially more believable when role boundaries remain incredibly explicit.
Review queues genuinely matter because they make autonomy feel safely operational rather than confusingly theatrical.
An optional model enrichment layer serves as a vastly superior onboarding path safely compared to forcing mandatory API keys from day one.
Above all, a highly small but fully inspectable orchestration engine proves decisively more useful than a large opaque one.
Top comments (0)