May 14, 2026
You know that exact moment when Claude marks a task "complete" but your browser automation is still clicking through screens doing god knows what? I burned three days tracking down why my Claude orchestrator kept reporting success while the actual Playwright instance was stuck in an infinite OAuth redirect loop. Split them up now - Claude writes the test scenarios and state checks, GPT-4 executes, and I run a third instance that just watches for drift between what Claude thinks happened vs what the browser console shows. Anyone else running a dedicated "verifier" instance to catch these silent failures?