Back to projects

Project

Multi-Agent Cognition Sandbox

Guardrailed simulation where LLM-controlled agents perceive, reason and act in a tile-based world.

pythonasynciopydanticllm agentspathfindingsimulation

Overview

A Python simulation sandbox where multiple LLM-controlled agents perceive a tile-based environment, reason about goals and execute actions through structured tool calls. Implements a reason–act–reflect loop with decisions parsed into Pydantic models, agent memory, async processing with asyncio, and terrain-weighted Dijkstra pathfinding.

Highlights

Loop
Reason · Act · Reflect
Pathfinding
Weighted Dijkstra
Runtime
Async