LABS & AUTONOMOUS AGENTS // QA HARNESS

Autonomous VLM Playtesting & Exploit-Hunting Agent

A custom Vision-Language Model agent harness that autonomous playtests, red-teams, and stresstests games via Playwright and Gemini multimodal vision. Discovers economic exploits, boundary violations, and state desynchronizations before human players encounter them.

HARNESS ARCHITECTURE & ENGINE

Execution Core: Node.js / Playwright Headless Browser Sandbox

Perception Model: Google Gemini Multimodal Vision (VLM)

Testing Modes: Compliant Playtester vs Adversarial Exploit Hunter

Watchdogs: Independent Invariant Checkers (Strict state diffs)

Clock Management: Dynamic clock pausing during VLM inference

BEYOND SCRIPTED TESTS

Traditional automated testing uses hardcoded button clicks that break on any UI shift. Our VLM agent physically views the canvas, understands game semantics, attempts irrational player actions ("deviate-then-serve"), and tests score validators against mathematical invariants.

Autonomous Watchdog Invariants

Economic Integrity Score increments strictly when a valid drink matches the independent recipe table, not game internal shortcuts.

Physical Mug Volume Limits Catches ingredient overflow exploits (e.g. attempting to pour 3+ shots into a 2-shot mug).

Monotonic Patience Decay Verifies customer patience timers never desynchronize or go into negative integers.