Vyuha / व्यूह
Agentic Sandbox — Game Theory x LLM Agents
Created on 15th February 2026
•
Vyuha / व्यूह
Agentic Sandbox — Game Theory x LLM Agents
The problem Vyuha / व्यूह solves
Multi-agent simulations today require writing code for every scenario — new rules, new behaviors, new interactions all mean new code.
Vyuha lets you build and control entire multi-agent worlds using just natural language.
Type "create two rival armies and make them fight" and watch 16 autonomous LLM agents spawn, strategize, and battle on a live 2D grid.
Mid-simulation, type "drop a mercenary in the center" or "a storm hits the battlefield" and the world changes instantly.
Each agent thinks independently — reading the world state, reasoning about its goals, and choosing actions (move, attack, cooperate,
trade, defend). A structured rules engine enforces consequences: agents die at 0 health, get penalized for breaking rules, and compete
using game-theory payoffs.
No code changes needed between scenarios. The same engine runs a war simulation, a resource economy, or a social experiment — all
controlled through English.
Challenges I ran into
The biggest challenge was drawing the line between engine code and God Mode responsibility. Early on, I kept hardcoding scenario-specific
logic (auto-death at 0hp, terrain rules) into the engine. But that breaks when the scenario changes. The solution: a two-layer rule
system — text rules that agents read and choose to follow, and structured rules that the system mechanically enforces. God Mode creates
both.
Agent intelligence was another battle. GPT-4o-mini agents couldn't distinguish teammates from enemies and kept attacking their own side.
Upgrading to GPT-4o fixed it, but at 15x the cost. The prompt engineering was critical — showing agents pre-computed distances, occupied
cells, and explicit range markers made the difference between chaos and coherent behavior.
State management on serverless was tricky. In-memory state broke on every hot reload and couldn't survive deployment. Migrated to Upstash
Redis (single JSON blob) which solved consistency across serverless function invocations.
Tracks Applied (1)
Hackathon Prizes
Technologies used