🏢 Inspiration
Inspired by the Stanford "Generative Agents" paper (Park et al., 2023), I wanted to build agents that don't just respond — they live. They wake up, remember their past, decide what to do, and act autonomously.
🤖 What it does
The Virtual Software House is a real-time simulation of a software studio where 4 AI agents — Alice (PM), Bob (Frontend Dev), Charlie (Backend Dev), and Diana (QA) — autonomously manage a real CRM project called Nova Logistics.
Every 15 seconds, each agent:
- Perceives their environment (zone, nearby colleagues, open tasks)
- Retrieves relevant memories using vector similarity + recency scoring
- Decides what to do via Gemini 3.5 Flash function calling
- Acts — moves, creates tasks, generates code, files bug reports, has conversations
- Stores new memories with embeddings for future retrieval
🛠️ How I built it
- Frontend: React + Firebase Realtime (Firestore subscriptions)
- Backend: Express server on Firebase App Hosting (Cloud Run) — Gemini API key lives here, never exposed to the browser
- AI: Gemini 3.5 Flash with function calling for structured agent decisions + gemini-embedding-001 for memory retrieval
- Database: Cloud Firestore (named database
virtual-software-house)
🧠 Challenges
The hardest part was making agents feel genuinely autonomous — not scripted. Getting the memory retrieval formula (recency × importance × relevance) to produce coherent, non-repetitive behavior took many iterations. Also implementing deduplication so agents don't spam the same tasks.
Log in or sign up for Devpost to join the conversation.