Vibe coding and agentic engineering are getting closer than I'd like

Willison's discomfort is the signal worth tracking: the convergence he describes isn't just aesthetic, it implies that the informal risk tolerance of vibe coding (move fast, accept hallucinations, iterate) is bleeding into agentic workflows that carry real production consequences.

This connects directly to two threads in recent coverage. The arXiv position paper from May 1st arguing that agentic orchestration should be Bayes-consistent identified the same structural problem from the opposite direction: current agent control layers are ad-hoc precisely because they inherited assumptions from exploratory LLM use rather than from principled engineering. Meanwhile, Willison's own iNaturalist project from that same week, built entirely on a phone via Claude Code, is a live demonstration of the blurring he now says concerns him. He was vibe coding an agentic tool in real time. The gap between those two moments, one practical and celebratory, one cautionary, is where the actual tension lives for practitioners.

Watch whether tooling vendors like OpenAI (Codex) or Anthropic introduce explicit guardrails or workflow-stage labeling that formally separates exploratory from production-grade agentic use within the next two quarters. If they don't, Willison's concern becomes a default industry posture rather than a correctable design choice.