Multi-Agent Systems for Instructional Design

Created: 2026-05-08 | Tags: agentic-aiai-literacyhuman-in-the-loopk-12pedagogical-safety
📄 Full text: arXiv:2508.16659 · local
Embedding the Knowledge–Learning–Instruction (KLI) framework into multi-agent systems to act as sophisticated instructional designers for K-12 educators.

The Prompting Gap in K-12 Education

Wang et al. (2025) address a critical problem: 60% of K-12 teachers now use AI for teaching tasks, but most lack the expertise to prompt effectively for high-quality instructional materials.

The Problem

The Solution

Shift pedagogical expertise from user's prompt to LLM's internal architecture by embedding learning sciences principles directly into multi-agent systems.

Three Systems Compared

System Architecture KLI Integration Teacher Preference
SAS (Single-Agent) Baseline: simple "Generate learning activity" prompt None (naive) Lowest: "needs lots of editing"
MAS-Roles 5-agent sequential pipeline (KC → Learning → Instruction → Design → Feedback) Embedded in agent roles Middle: "generic, impersonal"
MAS-CMD 3-agent collaborative with conquer-merge discussion (Behaviorist, Constructivist, Aesthetic personas) Collaborative discussion Highest: "fantastic, creative, classroom-ready"

Key Findings

Quantitative Results (QM K-12 Rubric)

Qualitative Results (20 Practicing Teachers)

MAS-CMD praised for:

Common critique: Pacing issues ("a lot to unpack," needs to be "broken up into some days")

Efficiency Trade-off

System Time (s) Tokens (k) Requests
SAS 25 ± 16 ~3.4 1
MAS-Roles ~75 ~24 5
MAS-CMD 272 ± 168 ~71.6 13

Finding: Highest quality requires ~11× more tokens and ~10× more time than baseline.

Knowledge–Learning–Instruction (KLI) Framework

The KLI framework (Koedinger et al., 2012) aligns three elements: 1. Knowledge Components (KCs): Facts, concepts, principles, procedures 2. Learning Processes: Memory/fluency, induction/refinement, understanding/sense-making 3. Instructional Principles: Spacing, testing, worked examples, prompted self-explanation

Embedding KLI into MAS:

Connection to Existing Work

vs. Principled AI Education

vs. Agentic Workflows

vs. TeachBench

Implications for AI in Education

For Tool Developers

For K-12 Educators

For Researchers

Related Pages

Sources