mabl's engineering team describes how they scaled their verification layer to match a 4.5x increase in PR throughput (179 to 810 PRs/month) driven by AI coding agents. Their stack includes seven layers from local TypeScript checks to 24/7 production monitoring, using two completely separate mabl stacks (prod and dev) that test each other to avoid shared fate. Key practices include: scoped test routing via Claude skills, automatic Claude Opus PR reviews with on-demand Codex second opinions, an MCP server that feeds deployment test results directly into coding assistants, and treating the test suite as a maintained product with composable flows, flake containment hierarchies, and self-describing plan names. 31% of PRs now flow through agent-only review. The core insight is that at AI-native throughput, the verification layer becomes the system's organizing principle, not a CI detail.
