GENESIS
Autonomous Startup Gauntlet
Training and evaluating LLMs on long-horizon, multi-agent coordination under uncertainty (the real reasons startups fail: planning, communication, morale/runway tradeoffs, and ethical crises).
An OpenEnv-compatible, stateful startup simulation where five role-specialized LLM “co-founders” run a B2B SaaS company via 42 MCP tools. Includes partial observability per role, an adversarial MarketMaker curriculum, personal/ethical crisis scenarios, pivot mechanics, and an 11-term composable reward with a GRPO (TRL + Unsloth) training pipeline and reproducible evaluation artifacts.
- Python
- FastAPI
- FastMCP
- OpenEnv
- MCP (JSON-RPC)
- Uvicorn
- GRPO
- TRL
- Unsloth
- Multi-agent
- Reward modeling
- Hugging Face Spaces


