Nova Agent Studio (NAS)

Inspiration

We wanted to build an agent that feels practical—not just a chatbot. Our goal was to show how Amazon Nova can power a “do the work” assistant that can plan, pull relevant context, and deliver a useful final draft in a clean chat UI.

What it does

Nova Agent is a chat-style web app where you give a goal (and optional context). The agent:

creates a plan when needed
retrieves relevant background context from a small knowledge base
writes a structured final answer (drafts, summaries, scripts, README content, etc.) It also supports - multi-turn conversation by including recent chat history in each request.

How we built it

Frontend: React + Vite + Tailwind chat UI
Backend: Node.js + Express (TypeScript)
Model: Amazon Nova 2 Lite via Amazon Bedrock (using an inference profile)
Agent loop: the model returns a JSON “decision” for either a tool call (makePlan, retrieve, writeDraft) or a final output, and the server executes tools and feeds results back until a final answer is produced
Deployment: Netlify hosts the frontend; Render hosts the backend
Security: requests are protected with an API key. The browser calls a Netlify Function proxy, which injects the key server-side so it’s not exposed in the frontend.

Challenges we ran into

Ensuring the model reliably followed a strict JSON-only contract for tool decisions
Handling edge cases where the model output included extra text/formatting
Deploying securely without exposing credentials or API keys in the browser

Accomplishments that we're proud of

What we learned

A polished chat UI with multi-turn memory and markdown rendering
Practical deployment/security tradeoffs for public demos (proxying + env vars)

What's next for Nova Agent Studio Agentic App (NASAA)

Add real document ingestion (upload PDFs/URLs) and semantic retrieval
Add user authentication + per-user rate limits
Streaming responses + better progress UI during tool loops
Persisted conversation history and more tools (web search, writing styles, exports)