A developer built an agentic AI harness to automate game playtesting, exposing a practical blueprint for AI-driven QA loops.
Jeff Schomay published a detailed blog post describing how he built an agentic test harness that lets an LLM autonomously play his game. The system feeds game state to the AI, receives actions, and loops — effectively automating playtesting without human involvement. The post gained 121 upvotes on Hacker News, signaling broad developer interest. No specific commercial tool was released; this is an open architectural blueprint with working code.
This is a working proof-of-concept for replacing manual playtesting with an agent that reads game state, generates actions, and iterates — a loop pattern directly transferable to any stateful app or simulation, not just games. The architecture is straightforward: serialize state to text, prompt an LLM to choose an action, execute it, repeat. The key insight is that LLMs can act as surprisingly capable exploratory testers when given well-structured state representations. This is essentially fuzzing with semantic intent.
Adapt Schomay's state-serialization loop to your own stateful app this week: pick your highest-complexity user flow, serialize the state to JSON, and prompt Claude or GPT-4o to choose the next action — measure how many edge cases it surfaces vs your existing test suite.
Suggested exercise · not independently tested. Access requirements and costs may vary.
Open Claude.ai and start a new conversation
Paste this prompt: 'You are a QA agent testing a web app. Here is the current app state as JSON: {"user": "logged_in", "cart_items": 3, "promo_applied": false, "checkout_step": "shipping"}. List the 5 most likely user actions to take next, ordered by likelihood of exposing a bug. For each, describe what edge case it targets.'
Review the ranked action list and compare against your existing test cases
Tags