The Design Space of LLM-Based AI Coding Assistants [pdf] (lau.ucsd.edu)

Researchers at UC San Diego released a comprehensive design analysis of 90 LLM-based AI coding assistants (58 industry products, 32 academic prototypes) and distilled their findings into a ten‑dimensional design space. The paper catalogs UI modalities, system inputs, capabilities, and outputs—grouped as user interface (development environment, user actions, initiative), system inputs (input format, semantic context, personalization), capabilities (autonomy, system actions), and outputs (output format, explainability). By reviewing papers, docs, demos and tooling from 2021–mid‑2025, the authors identify three UI eras (tab autocomplete, chat, and agent-based interfaces), show that industry products are converging on speed and a common feature set while academic projects diversify toward scaffolding and metacognitive support, and define six user personas (professional engineers; HCI researchers/hobbyists; UX designers; conversational programmers; data scientists; students). This work matters because it creates a shared vocabulary and map for tool builders, researchers, and evaluators to reason about tradeoffs (e.g., autonomy vs. explainability, inline autocomplete vs. multi‑turn chat vs. autonomous agents). The design space makes concrete how differing user needs should drive interface and capability choices, highlights gaps (pedagogical scaffolding, validation and safety), and suggests directions for benchmark design, UX studies, and responsible deployment—helping the community move from hype and isolated prototypes toward systematic, user‑centered engineering of coding assistants.