Choose Your Own Adventure: AI-Assisted Programming with EvoGraph

Why Read This from an HCI Perspective

This article reframes AI coding tools not as ‘code generators,’ but as interfaces that help users follow and understand the work together. Rather than focusing only on whether the answers are correct, it shows why it matters to design for safe experimentation—so users can try multiple approaches, bring prior context back into view, and backtrack when they take the wrong path. For HCI/UX practitioners and researchers, it’s a piece that prompts reflection on what actually changes the real usage experience when adding AI capabilities—not just what feels convenient.

CIT's Commentary

The core of EvoGraph isn’t about adding more ‘smartness’ to the LLM. It’s about enabling people and AI to leave shared traces so the work can be understood together. In particular, the problem is that linear chat doesn’t match what programming really looks like: iterating across multiple branches, experimenting, and revisiting earlier points. This also carries over directly to product design. As features grow, graphs can become more useful, but they can also become easier to overcomplicate. Automatic logging is convenient, but users must be clear about when to place checkpoints and where to intervene. In development environments where safety is critical, ‘ability to roll back’ and ‘traceability of causes’ shouldn’t be optional add-ons—they should be fundamental interaction primitives.

Questions to Consider While Reading

• Q.In real products, how often—and at what moments—does a graph-based development history provide the greatest value?
• Q.Between automatically recording many checkpoints and letting users place checkpoints themselves, what balance is most appropriate?
• Q.When showing code generated by AI versus code edited by humans, where is the boundary between improving developers’ understanding and making the screen too complex?