verify

Coding agents say "done" before anyone has opened the app. Told to test on web and Android, they test one platform or neither, and parallel agents fight over the same simulators. verify fixes that with one command. When an agent finishes a change, it runs verify "<what changed>", or /verify through the Claude Code skill. The working tree goes to a control VM on Vultr Cloud Compute. An independent verifier on Vultr Serverless Inference (glm-5.3) reads the diff, writes its own checks, and drives the app in a fresh sandbox on a Vultr VX1 worker: Playwright in Chromium for web, an Android 14 emulator with KVM and Maestro for mobile. It returns a verdict with screenshots, a recording, network and memory facts, and steps to reproduce. The agent fixes the bug and runs verify again before it reports back. Every run gets its own container and private network, no secrets, CPU and memory caps, and a proxy that allows the package registry during install and nothing after, and is destroyed afterwards. The demo shows an install script blocked from sending data out and a memory leak killed at the cap. When the queue backs up, the control VM adds workers through the Vultr API and removes them when idle. So far: 47 platform runs, a median of 49 s on web and 4 min 46 s on Android, about $3 of Vultr charges in total. Today: web apps, and Expo or React Native apps on Android. Later: iOS, more devices and browsers, API checks, migrations, CLI tools and proof on pull requests.