Show HN: AI Game Playtester
Ziva's Playtest Agent automates game QA by using AI agents that play games end-to-end, replacing costly human playtesting that costs studios ~$1.7B annually The agent reads game code, injects temporary hooks and logging, then drives the game turn-by-turn with real keyboard/mouse input while measuring outcomes via screenshots and custom signals It pauses the game between LLM decisions (since LLMs aren't real-time), runs dozens of parallel instances, and produces a pass/fail verdict with recorded
Analysis
TL;DR
- Ziva's Playtest Agent automates game QA by using AI agents that play games end-to-end, replacing costly human playtesting that costs studios ~$1.7B annually
- The agent reads game code, injects temporary hooks and logging, then drives the game turn-by-turn with real keyboard/mouse input while measuring outcomes via screenshots and custom signals
- It pauses the game between LLM decisions (since LLMs aren't real-time), runs dozens of parallel instances, and produces a pass/fail verdict with recorded video for each test
- Currently supports Godot 4.2+ natively, with Unity via an editor bridge and Unreal/custom engines through Enterprise; PC, mobile, and console are all supported
- Runs fully locally with only zero-data-retention LLM calls for reasoning and image analysis, keeping project files on the developer's machine
Why It Matters
AI-driven playtesting addresses a massive bottleneck in game development: functional QA that traditionally requires recruiting humans, waiting days for results, and spending millions annually. By automating the "does it work" layer, studios can run comprehensive test suites on every build in minutes, freeing human testers to focus on experience, feel, and design feedback rather than broken buttons and edge cases.
Technical Details
- Code-first analysis: The agent reads the game's source code to understand input handling, game logic, and where to inject temporary hooks and logging for measurement
- Paused execution model: Since LLMs cannot operate in real-time, the game is launched in a sandbox and frozen between turns, giving the model unlimited time to reason before sending the next input
- Multi-modal feedback loop: Inputs are sent as real keyboard/mouse/gesture commands; outputs are measured through a combination of screenshots, custom hooks that respond to LLM queries, and added logging
- Parallel execution: Dozens of instances run simultaneously, enabling full test suites to complete in minutes rather than hours or days
- Verdict generation: After completion, the AI reviews all collected data holistically to produce a pass/fail result and saves a playback video for human review
- Local-first architecture: The game runs entirely on the developer's machine; only anonymized reasoning and image analysis requests go to external LLMs with zero data retention
Industry Insight
- The $1.7B annual playtesting spend represents a significant cost center that AI automation can compress dramatically, especially for indie and mid-size studios that lack dedicated QA teams
- This tool shifts the human playtester role from functional verification to experience evaluation, suggesting a new workflow where AI catches bugs first and humans assess fun and polish second
- The engine-agnostic approach (Godot, Unity, Unreal, custom) signals that AI playtesting will become a standard layer in the game dev pipeline, similar to how CI/CD became ubiquitous in software development
Disclaimer: The above content is generated by AI and is for reference only.