AI Agents with persistent memory, tool access, and specialist skills can now execute multi-step reasoning across the entire research pipeline, from idea to submission. This is not merely the fourth wave of research automation — following statistical computation, digital trace data, and machine learning — but the first to automate reasoning itself, and it raises a question that belongs to the sociology of science tradition: what happens to the social organization of knowledge production when machines can perform tasks that previously required years of specialized training? This talk situates vibe researching as a sociological rather than purely technical question. AI agents disrupt Mertonian norms in distinct ways — complicating communalism, challenging universalism, blurring disinterestedness, and straining organized skepticism — and the delegation boundary is cognitive, not sequential, separating codifiable execution (delegate) from tacit judgment, theoretical originality, and field knowledge (protect). A generation–verification asymmetry further structures the human–AI relationship: AI compresses production time by orders of magnitude while verification time remains largely unchanged, with consequences unfolding at individual, institutional, and epistemic levels. Using Open Scholar Skills — a Claude Code plugin the speaker developed covering research workflow from idea formation, peer-review simulation, to replication — as a live case study, the talk closes with the stratification, pedagogical, and normative stakes for the discipline.