Reality Fork

Agent-powered decision simulation. Give it a life decision, and it generates parallel "what-if" timelines so you can see how each path might unfold.

• Adaptive scope: a conversation decision gets simulated in days; a career decision gets simulated in months or years. The planner agent picks the horizon.
• Parallel multi-agent simulation: up to four specialized agents (financial, career, psychological, events) score each fork step-by-step, then a narrator agent stitches them into a cohesive story.
• Side-by-side comparison: every fork gets its own timeline, plus line charts comparing forks on each dimension.

Built for the UMBC Hackathon 2026.

Tech

• Next.js 16 (App Router, TypeScript)
• Tailwind CSS 4 + custom UI primitives
• Recharts for metric charts
• Zod 4 for schema contracts
• LLM: pluggable (with a shared rate-limited wrapper that throttles requests-per-minute and backs off on 429s using the server's retry hint)
  • MiniMax (default) — M2.7-highspeed via the Anthropic-compatible endpoint. Structured output is enforced via forced tool use.
  • Google Gemini — grammar-enforced responseSchema.
  • Ollama Cloud — free, format hint only (less reliable for strict JSON).

Run locally

pnpm install
cp .env.example .env.local
# edit .env.local and paste MINIMAX_API_KEY (or switch LLM_PROVIDER)
pnpm dev

Open http://localhost:3000.

A real simulation fires ~11 LLM calls (planner + forks × dimensions + narrators) and takes roughly 2–5 minutes end to end. The loading screen is not frozen.

Switching provider / model

Set env vars in .env.local:

# default — MiniMax M2.7-highspeed (Anthropic-compatible)
LLM_PROVIDER=minimax
MINIMAX_API_KEY=sk-...
# LLM_MODEL=MiniMax-M2.7-highspeed
# MINIMAX_BASE_URL=https://api.minimax.io/anthropic

# or: Gemini (free tier)
LLM_PROVIDER=gemini
GEMINI_API_KEY=...
# LLM_MODEL=gemini-2.5-flash-lite

# or: Ollama Cloud (free)
LLM_PROVIDER=ollama
OLLAMA_API_KEY=...
# LLM_MODEL=qwen3.5:397b

# optional: override the per-minute request cap (defaults: minimax=60,
# gemini=5, ollama=60)
# LLM_RPM=30

The provider interface is in lib/llm/types.ts. Add a new provider by implementing LlmProvider.generateStructured and wiring it in lib/llm/index.ts. Every provider is automatically wrapped in RateLimitedProvider.

Architecture

User decision + context
  ↓
[Planner agent] → horizon, granularity, dimensions, forks
  ↓
for each fork (parallel):
  ├─ [Financial agent]      ┐
  ├─ [Career agent]         │ parallel structured-output calls
  ├─ [Psychological agent]  │ (dimensions list is adaptive)
  └─ [Events agent]         ┘
  ↓
[Narrator agent] → cohesive step-by-step timeline
  ↓
Side-by-side UI + metric charts

All agent I/O is typed via shared Zod schemas in lib/schemas.ts. Orchestration lives in lib/orchestrator.ts and uses nested Promise.all — the planner decides which dimensions matter for a given decision, so some runs only spin up 2–3 dimensional agents per fork.

Deploy

vercel
# add MINIMAX_API_KEY (and LLM_PROVIDER if not default) in the Vercel dashboard

The /api/simulate route sets maxDuration = 60. A full simulation can exceed that — use Vercel Pro for a longer function timeout, or move to a streaming response that emits forks incrementally.