Nova Forge - Agentic System

Inspiration

We watched developers spend hours on boilerplate — scaffolding projects, wiring APIs, debugging CSS. What if you could just describe what you want and an AI builds the whole thing? Amazon Nova's 3-tier model family (Lite/Pro/Premier) gave us the perfect playground: one framework, three capability levels, real production output.

What it does

Nova Forge turns a single English prompt into a deployed application. Type "build a tower defense game with 6 tower types" and Nova agents plan the architecture, decompose it into tasks, execute them in parallel waves, run adversarial quality reviews, and deploy to a live URL. 9 playable game demos — from simple arcade to pseudo-3D racing — all built by Nova with zero human code.

How we built it

19 sprints, ~30,000 lines of pure Python over 12 days. The core is a tool-use agent loop that works with any LLM supporting function calling. We built 14 agent tools (read, write, edit, replace_lines, bash, grep, glob), 11 team formations, a 3-tier prompt system optimized per context window, and a 6-level autonomy system. 1,670 tests verify everything. All 3 Nova models score S-tier (100%) on our benchmark suite.

Challenges we ran into

Nova models would describe code instead of writing it — we traced this to missing tool-use instructions in the agent prompts. Files landed in src/ instead of the project root. The decomposer stripped technical detail from tasks, causing models to hallucinate frameworks. The verify phase triggered too early, killing multi-file builds. Each bug required tracing through the full pipeline to find the root cause.

Accomplishments that we're proud of

Nova Pro debugged its own game — found and fixed 5 critical bugs in 9 turns, 26 seconds, zero manual intervention. The replace_lines tool we built enables structural code edits that string-matching can't handle. Every demo card on our site has an expandable "proof of work" showing the exact prompt, model, build stats, and task breakdown. All verifiable, all real.

What we learned

Small context windows need fundamentally different prompts, not just shorter ones. Preserving the user's original goal verbatim through every pipeline stage matters more than elegant summarization. And convergence detection needs to respect the creative phase — don't kill writes just because the agent paused to think.

What's next for Nova Forge

Open-source release with pip install. Real-time streaming UI so you watch agents work. Multi-project orchestration for microservice architectures. And a community demo gallery where anyone can submit what Nova built for them.

Built With

• amazon
• amazon-nova
• bash
• bedrock
• boto3
• cli
• github
• python
• typescript