this is pure f*cking insane Teams blame Claude Sonnet 5.5 and GPT-6 for crashes, but benchmark telemetry proves 92% of reasoning failures are unmonitored trajectory drift When autonomous agent workflows break in production, developers reflexively tweak prompt templates or dump