Agentic Sequential Reading: How Scaffolded Incremental Processing Shapes AI Literary Engagement

We present agentic sequential reading, a three-phase pipeline in which an AI agent reads novels through incremental chunked processing with structured per-chunk reflections, cumulative context tracking, and a final synthesis. Drawing on a dataset of 41 complete Discworld novel reads, we conduct controlled experiments comparing sequential reading against single-pass baselines across three conditions: no context, series-aware context, and full sequential processing. Our central finding is that the method effect on evaluative ratings decomposes into two distinct components: a context effect, where cumulative series knowledge sharpens evaluative discrimination in both directions (raising ratings for series-embedded books while lowering them for books representing quality regression), and a process effect, where incremental engagement adds value for books whose quality lies in the reading experience itself. Sequential processing also produces structurally different output containing temporal markers, predictions, and revisions that single-pass reviews cannot generate by design. Cross-model experiments show that sequential scaffolding partially compensates for model capability differences, functioning as performance augmentation rather than capability transfer. Grounded in reader-response theory (Rosenblatt, Iser, Jauss) and cognitive science (Kintsch, Zwaan), these findings suggest that structured reading processes produce measurably different AI literary engagement, with implications for how we design and evaluate AI systems that engage with long-form text.