Generative AI (GenAI) is increasingly used in qualitative research, but methodological debate has largely centred on conversational chatbots. We review this literature and organise recurring concerns into ten methodological challenges, alongside five epistemic risks identified in prior work. We then use an illustrative multi-agent workflow for reflexive thematic analysis, with ChatGPT as a point of comparison to examine how these methodological challenges and epistemic risks are reconfigured under a different configuration. The analysis shows that agentic design can constrain challenges associated with data handling, model control, and transparency, while non-replicability, hallucination, and evaluation uncertainty are shifted into more inspectable parts of the workflow. Bias and environmental burden remain largely unresolved. The workflow does not remove the underlying epistemic limits of LLM-based analysis or the need for reflexive researcher judgement. The study’s primary contribution is a risk-reconfiguration account showing how GenAI configuration redistributes methodological dependence while interpretive responsibility remains with the researcher.