The design space of AI coding tools (austinhenley.com)

Researchers Sam Lau and Philip Guo (UC San Diego) surveyed 90 AI coding assistants from 2021–2025 (58 industry, 32 academic) and mapped the design space into four broad categories—user interface, system inputs, capabilities, and outputs—subdivided into 10 concrete dimensions: development environment, user actions, initiative, input format, semantic context, personalization, autonomy, system actions, output format, and explainability. They trace three UI eras (autocomplete inline suggestions, multi-turn chat, and agent-style tools that read/write repos and run commands) and show industry converging on a polished bundle (inline + chat + agents) prioritizing latency and breadth, while academia explores novel interactions like reflective scaffolding, multimodal sketch-to-code, proactive secondary cursors, and improved explainability. The paper also defines six user personas and maps how design choices trade off for each. The significance is practical: this taxonomy clarifies concrete trade-offs—flow vs. context (inline vs. chat), control vs. convenience (user-driven vs. autonomous agents), personalization vs. cognitive overload—and highlights risks such as increased failure surface with more autonomy. For builders and evaluators it offers a framework to choose metrics (latency, correctness, explainability, integration, safety) aligned to target personas and workflows. The authors argue for combining industry’s production rigor with academia’s interaction innovation to move assistants from “it writes code” novelty toward trustworthy, workflow-aware tools.