Willison identifies November 2023—not 2025—as the true inflection point when AI coding agents became reliably production-ready, enabling him to write 95% of his code from a phone despite mental exhaustion by mid-morning. He distinguishes casual 'vibe coding' from disciplined 'agentic engineering', highlighting three core patterns: red/green TDD (where agents write and run tests autonomously), curated templates for consistency, and 'hoarding'—maintaining public, searchable repositories of past work to guide future AI output. The 'dark factory' model—fully autonomous build-test-deploy pipelines—is already emerging, with AI performing rigorous security testing like Firefox vulnerability discovery. Yet critical risks remain: prompt injection is still unsolved, and the 'lethal trifecta' (private data + adversarial input + exfiltration) poses systemic danger. Willison warns that 97% reliability is dangerously insufficient, echoing the Challenger disaster’s normalization of deviance. Meanwhile, human value is shifting toward idea synthesis, judgment, and cross-domain integration—not coding itself. He predicts half of engineers will write 95% AI-generated code by end-2026, demanding a pivot toward architecture, validation, and agency—not just automation.