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 studied 93 middle school students who built their own AI agents using CocoFlow, a no-code platform. Researchers measured computational thinking before and after, analyzed behavioral logs, and conducted interviews. The key finding: the largest gains went to students with middling initial computational thinking—not the strongest or weakest. The effect size was η² = 0.55, described as an optimal development zone effect. Middle-performing students showed adaptive expertise, flexibly moving between abstraction and concreteness. High-performing students tended toward over-engineering, writing complex nested logic where simple conditionals sufficed, while low performers struggled with task decomposition, unable to break a broad goal like building a customer service agent into steps. Abstraction improved by 0.71 effect size and algorithmic thinking by 0.70; participation in iterative testing predicted self-efficacy gains (β = 0.20, p = 0.05). Limitations include the short five-day duration and the possibility that no-code interfaces mask core computational thinking challenges needed for future text-based coding.

Designing AI agents requires computational thinking: you have to break an agent's behavior down into states, conditions, and action sequences—an exercise in abstraction and algorithmic thinking. But a question follows: do students with strong computational thinking keep gaining more, while those with weak skills fail to learn at all?

Sun, Xin, Niu, Li, Huang, and Chen addressed this with a five-day AI agent creation workshop. 93 middle school students built their own AI agents using CocoFlow, a no-code platform. Computational thinking was measured with pre- and post-tests, supplemented by behavioral logs and interviews.

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

Key findings

Middle-level students gained the most. Not the strongest, and not the weakest baseline group. The effect size was η² = 0.55—an "optimal development zone" effect.
Middle group: adaptive expertise. These students knew when to abstract and when to stay concrete, without rigidity.
High group: over-engineering. Students with strong computational thinking tended to write complex multi-level nested logic for problems that simple conditional statements could solve.
Low group: stuck on task decomposition. They did not know where to start—for example, how to break "build a customer service agent" into steps like defining user intents, designing reply templates, and configuring dialogue flows.

Specific numbers

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

Open questions

The workshop lasted only five days—what are the long-term effects?
CocoFlow's drag-and-drop interface shields students from programming complexity. But that hidden complexity may be exactly the core challenge of computational thinking. If students later need to implement agents in code, will these skills transfer?

---

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.