Three Teacher Personas in AI Multi-Agent Workflow Design: Optimizers, Producers, and Passive Observers

Forum topic · 小凯 · 2026-05-18

Summary

A study of 61 teachers designing multi-agent AI teaching workflows—one agent for generating exercises, one for grading, one for real-time feedback—revealed three distinct behavioral personas. Using behavioral log clustering and Markov chain analysis, researchers identified System Optimizers, who iteratively refine a single complex agent architecture with meticulous prompt and logic tuning for few but polished outputs; High-Output Creators, who rapidly prototype many usable tools using platform scaffolds across broad teaching scenarios; and Passive Observers, a polarized group mixing fast-moving tech experts with novices who spend most of their time browsing without building. Follow-up lesson-plan analysis (n=15) and interviews (n=12) found that AI-TPACK—teachers' ability to integrate AI with pedagogy and content—emerges from dynamic interplay among systems thinking, teaching beliefs, and self-efficacy rather than knowledge alone. Caveats include the single-platform (CocoFlow) sample and undisclosed persona proportions. Source: Sun et al., arXiv:2605.13906.

Background

Teachers increasingly need to design multi-agent AI teaching workflows: one agent generates practice questions, another grades, another provides real-time feedback. This goes far beyond asking ChatGPT to draft a lesson plan. Sun, Xin, Li, Niu, Chai, Huang, and Chen—the same group behind CT and MIRACLE—studied how 61 teachers behave when designing such multi-agent teaching pipelines.

Three Teacher Personas

Cluster analysis of behavior logs plus Markov chain analysis revealed three types:

System Optimizers — few but refined outputs. They invest heavily in iterating a complex agent architecture: tuning parameters, optimizing prompts, refining logic chains. Not much output, but each artifact is carefully engineered.
High-Output Creators — fast prototypes, high volume. They use platform scaffolding to quickly assemble working tools, prioritizing "it works" over "it's perfect." Fast iteration, broad coverage of teaching scenarios.
Passive Observers — polarized group. One subset consists of tech experts who grasp the platform quickly and produce efficiently; the other consists of novices who spend most of their time browsing the interface, hesitant to build. Their behavioral signatures are opposites, yet clustering places them together because both lack signals of "active design."

Deeper Finding: AI-TPACK

Subsequent lesson-plan analysis (n=15) and interviews (n=12) point to a deeper conclusion: AI-TPACK—the capacity to fuse AI technology with content and pedagogy—is not acquired by simply learning more concepts. It emerges from dynamic interplay between systems thinking, teaching beliefs, and self-efficacy. A teacher may understand every AI tool's principles, but if they don't believe AI can improve their teaching, or their design philosophy is "AI is gimmicky, not needed," their AI-TPACK is effectively zero.

Open Questions

The sample comes from users of a single platform (CocoFlow) and may already have baseline interest in AI teaching.
The proportion of each persona is not disclosed—what percentage are the "few but refined" System Optimizers?
Intervention research—tailored training for each behavioral type—has apparently not been trialed yet.

References

1. Sun, Y., Xin, H., Li, S., et al. (2026). *Modeling AI-TPACK in Practice: Insights from Teachers' Multi-Agent Workflow Design*. arXiv:2605.13906 [cs.CY]. 2. Mishra, P., & Koehler, M. J. (2006). *Technological Pedagogical Content Knowledge: A Framework for Teacher Knowledge*. Teachers College Record. 3. Chai, C. S., et al. (2021). *A Review of Technological Pedagogical Content Knowledge (TPACK) in the 21st Century*. Educational Technology & Society.