Choose Your Own Adventure: Non-Linear AI-Assisted Programming with EvoGraph
The IDE plugin automatically records AI interactions as a graph, letting developers compare and merge different code paths.
A team of researchers from Carnegie Mellon University and the University of British Columbia has published a paper introducing EvoGraph, a novel IDE plugin designed to fundamentally reshape how developers interact with AI coding assistants like GitHub Copilot. The core innovation addresses a critical flaw in current tools: their linear, chat-based nature clashes with the inherently iterative and experimental process of software development. EvoGraph automatically captures every AI prompt, response, and subsequent code change, weaving them into a lightweight, visual development graph. This transforms a confusing sequence of prompts into a navigable map of the programmer's thought process and the AI's contributions.
In a user study with 20 participants, EvoGraph proved highly effective. Developers reported a lower cognitive load compared to traditional linear interfaces, as the graph provided a clear visual history. More importantly, it enabled "safe exploration"—users could freely ask an AI for multiple solutions to a problem, visualize them as separate branches in the graph, and then directly compare, test, and even merge the best parts from different AI-generated code variants. This non-linear workflow directly tackles the common struggle of managing prompting sequences and tracing changes, turning AI from a one-shot code generator into a collaborative partner in a true exploration of the solution space.
- EvoGraph is an IDE plugin that visualizes AI coding sessions as an interactive, branching graph instead of a linear chat.
- It automatically records all AI interactions and code changes, allowing developers to compare, merge, and revisit different solution paths.
- A study with 20 developers found it reduced cognitive load and supported safer, more efficient iteration on AI-generated code.
Why It Matters
It moves AI coding beyond simple prompt-and-copy, enabling systematic exploration of alternatives and making AI a true collaborator in complex software design.