Human–Multi-Agent Teaming for Higher-Order Thinking

The problem

Single-agent chat tutors tend to hand students finished answers, which flattens the higher-order thinking that data science education is meant to build.

Research question

Can a coordinated team of narrowly-scoped LLM agents scaffold a student's reasoning in AI and data science, rather than replace it, and can that be observed in how learners interact?

System design

A learning environment in which a student works with role-specialised LLM agents. Orchestration is built in LangGraph behind a FastAPI backend, and agents are constrained by role-restricted prompts to scaffold rather than hand over complete solutions. SHAP and LIME ground the explainability agent in actual model behaviour.

Agent architecture

Nine specialised agents sit behind a tutor-facing entry point. Data-preparation roles (dataset exploration, cleaning, feature engineering) feed an ML advisor, a critic challenges the proposed approach, and explainability and visualisation agents close the loop.

Method

Interaction telemetry was logged per session and analysed turn by turn, classifying student turns and tracking how the share of higher-order-thinking interactions changes as a session progresses.

Higher-order-thinking interactions rose from 5.6% at the start of a session to 16.5% by the sixth turn.

The 5.6% → 16.5% figure describes how interactions change within a session. Telemetry is observational, so on its own it does not establish that the agent team caused the change.