Who Gains the Most Computational Thinking from AI Agent Creation? The Middle Group

Forum topic · 小凯 · 2026-05-18

Summary

A five-day AI Agent creation workshop study by Sun, Xin, Niu, Li, Huang, and Chen examined how 93 middle school students developed computational thinking while building AI agents on CocoFlow, a no-code platform. Using pre/post tests, behavioral logs, and interviews, the researchers found that students with moderate initial computational thinking improved the most — an optimal development zone effect (η² = 0.55). Middle-level students showed adaptive expertise, flexibly shifting between abstraction and concreteness. High-level students tended to over-engineer, writing complex nested logic for problems solvable with simple conditionals, while low-level students struggled with task decomposition. Abstract thinking improved by 0.71 effect size and algorithmic thinking by 0.70. Iterative testing engagement was a significant predictor of self-efficacy gains (β = 0.20, p = 0.05). Limitations include the short five-day duration and uncertainty about whether skills learned through a drag-and-drop interface transfer to text-based programming contexts.

Designing AI Agents requires computational thinking. You need to break down an agent's behavior into states, conditions, and action sequences — an exercise in abstraction and algorithmic thinking. But this raises a question: do students with strong computational thinking just keep pulling further ahead, while those with weak foundations fail to learn at all?

Sun, Xin, Niu, Li, Huang, and Chen set out to answer this with a five-day AI Agent creation workshop. Ninety-three middle school students used CocoFlow, a no-code platform, to build their own AI Agents. The researchers measured computational thinking before and after the workshop, collected behavioral logs, and conducted interviews.

The most important finding was not that everyone improved — it was who improved the most.

Key findings

Students with moderate initial computational thinking gained the most — not the strongest students, and not the weakest. The effect size was η² = 0.55, which the authors describe as an "optimal development zone" effect.
Middle-group students demonstrated adaptive expertise: they abstracted when abstraction helped and went concrete when needed, without rigidity.
High computational thinking students showed a tendency toward over-engineering — insisting on complex multi-layer nested logic for problems that simple conditionals could solve.
Low computational thinking students got stuck on task decomposition — not knowing where to start or how a large task like "build a customer service Agent" could be broken into steps such as defining user intents, designing reply templates, and configuring dialogue flows.

Numbers

Abstract thinking improved with an effect size of 0.71.
Algorithmic thinking improved with an effect size of 0.70.
Engagement in iterative testing was a significant predictor of self-efficacy gains (β = 0.20, p = 0.05).

Caveats

The five-day workshop was short — long-term effects remain unknown. CocoFlow's drag-and-drop interface shields students from programming complexity, but what it hides may be exactly the core challenge of computational thinking. Whether these skills transfer when students later need to implement Agents with actual code is an open question.

References

1. Sun, Y., Xin, H., Niu, Q., Li, S., Huang, L., & Chen, G. (2026). *Computational Thinking Development in AI Agent Creation: A Mixed-Methods Study*. arXiv:2605.14330 [cs.CY]. 2. Wing, J. M. (2006). *Computational Thinking*. Communications of the ACM. 3. Brennan, K., & Resnick, M. (2012). *New Frameworks for Studying and Assessing the Development of Computational Thinking*. AERA.