Qwen3.6 model autonomously builds A* pathfinding in Java game after 12-hour testing marathon
A developer used Claude Code with a local Qwen3.6 model to create A* pathfinding entirely through autonomous testing.
Reddit user swagonflyyyy used Claude Code locally running an unspecified model to implement A* pathfinding for a test game. The model autonomously ran a testing suite for nearly 12 hours, monitoring logs, refactoring code, and relaunching the game in real-time, with occasional manual tests. After iterative attempts, the NPC can now climb over blocks, drop down, and navigate around gaps and obstacles smoothly — though it still crashes with index errors when the player is too far away, and the navigation works most of the time.
- Used Alibaba's Qwen3.6-27b-mtp-q8 model locally via Claude Code to code A* pathfinding in Java.
- Model built an autonomous testing suite that monitored logs, refactored code, and relaunched the game for 12 hours.
- Final NPC can climb, drop, and navigate obstacles smoothly, though occasional bugs persist at long distances.
Why It Matters
Demonstrates how local AI models can autonomously debug and refactor code over long iterations, reducing manual intervention in game dev.