"Trusting the Pipeline, Not the AI": Agency and Adoption in AI-Assisted Software Development

AI coding assistants have become integral to professional software development, yet how developers experience trust, maintain professional agency, and navigate adoption remains poorly understood. We address this gap through 26 semi-structured interviews with professional software developers across six countries, analyzed using reflexive thematic analysis grounded in established trust models. Our analysis yields three categories of contribution. First, we offer empirical grounding for trust laundering, a mechanism whereby developers route trust through verification infrastructure rather than placing it in AI itself. Our data reveal that this mechanism operates in layers, from functional testing through architectural review to design principle adherence. Second, we identify how developers mental models of AI co-occur with distinct trust calibration patterns, document persistent ambivalence as a trust state, and uncover an epistemological shift from learn-then-build to build-then-learn that restructures professional agency. Third, we document social adoption dynamics including shame, compliance theater, and management disconnect alongside concerns about a junior developer pipeline risk. Together, these findings suggest that trust, agency, and adoption are structurally intertwined, with implications for human-centered AI design, trust calibration, and sustainable organizational adoption.