This paper presents Knoweia, a dialogue-centered AI learning system designed to support computational social science education, where students must coordinate causal reasoning, research design, coding, and data analysis. Rather than treating large language models as answer generators, Knoweia functions as instructional infrastructure that organizes learning as a guided and stateful process.

We introduce the concept of Instructional Navigation Capacity (INC), defined as a learner’s ability to orient within a task space, interpret obstacles, and progress under uncertainty. INC is operationalized using process-trace data, including persistence after errors, recovery from repeated blockers, structured task progression, and patterns of help-seeking.

Knoweia implements this framework through a multi-agent architecture consisting of a learner-facing Companion Agent, a Roadmap Manager that tracks and updates task trajectories, a Memo Agent that records longitudinal learning patterns, and an expert layer for controlled escalation. The system is deployed in an ongoing course with 236 students from diverse disciplinary backgrounds.

We combine baseline survey data with fine-grained interaction traces to examine how dialogue-centered AI guidance relates to learning-process outcomes such as persistence, error recovery, and trajectory coherence. By linking conceptual framing, system design, and behavioral data, the paper contributes a process-oriented framework for evaluating human–AI learning systems in computational social science.