What is it about?

Many programmers can write code, but understanding how complex, professional software systems are designed is much harder. This deeper understanding is especially important today, when developers must review, adapt, and fix AI-generated code. However, there are far more learners than expert mentors, and most existing tools only help with narrow technical tasks like debugging. In this research, we introduce a framework called Process Management (PM) that helps intermediate programmers learn from real-world professional code in a structured way. Instead of simply reading code, learners are guided to compare examples, map patterns, reflect on design decisions, and uncover why certain techniques are used. We built and studied three PM-based tools for JavaScript, CSS, and Python. In our studies, learners using PM tools demonstrated more systematic exploration and achieved about three times more meaningful learning outcomes than those using standard tools. Toward the end of this work, we prototype AI agents that automate parts of the experimenter’s facilitation process, suggesting a path toward greater scalability.

Featured Image

Why is it important?

As software systems grow more complex and AI-generated code becomes commonplace, the bottleneck is no longer coding itself but the ability to deeply understand and evaluate architectures and systems. This work addresses an underexplored stage of professional growth: how intermediate programmers develop expert-level conceptual insight from real-world systems. By introducing and empirically validating a structured learning process that consistently improves higher-level understanding across domains, this research offers a scalable model for professional development and a foundation for AI-assisted systems that can help engineers build transferable expertise (which can meaningfully complement AI agents) rather than just produce more code (which AI agents can already do).

Perspectives

There are a few things that I think make this work unique and personally interesting: A) It's rare to find a paper that studies hundreds of hours of intermediate-level programmers making sense of professional systems. B) Understanding, engineering, and aligning systems to human needs, in my opinion, is the most valuable economic activity a human can do. As AI agents (Claude Code, Cursor, etc.) tackle more knowledge work domains, understanding systems, patterns, and architectures enough to steer agents toward human interests may be one of the few jobs we "humans in the loop" will have left. C) The way we approach the construction of automatable systems is fundamentally democratizing. "Taste" and "tribal knowledge" need not be so esoteric. AI is already learning from real-world codebases, and if AI can also help diverse humans learn from those systems, we can create symbiotic human-AI systems that deliver outcomes much more efficiently.

Gobi Dasu
Northwestern University

Read the Original

This page is a summary of: Process Management for Learning from Professional Source Code: Cultivating Experts in the Age of AI, ACM Transactions on Computing Education, February 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3796525.
You can read the full text:

Read

Contributors

The following have contributed to this page