A2UI vs MCP Apps: competing approaches to AI agent interfaces

On October 3, 2026, a technical post on V2EX compared two competing approaches to how AI agents present interfaces to users, framing the question as another round of the long-running low-code versus high-code debate in front-end development. The author's starting point is that most AI systems today are good at answering but not at interacting. Programmers, in this view, do not need much interaction with an AI, since what a programmer produces is essentially what the AI returns: code. Everyone else needs interaction far more.

The post argues that the most common interaction pattern so far is the AI asking a question and a person answering it, which it likens to the earliest days of internet BBS forums. The next stage, it suggests, should be the era of web applications. From there it lays out two mainstream routes. The first is Google's A2UI, a low-code approach in which the agent's client has the greatest possible control over the interaction page. The author calls his own position on this "very radical": low-code is a dead end for general-purpose scenarios, and the agent client is exactly such a general-purpose scenario, unless the agent targets a highly specialized business. The second route is MCP Apps from Anthropic and OpenAI, which mounts finished web applications inside the conversation.

The author says he prefers the MCP Apps approach, though the passage in which he explains why is cut off in the available material, leaving only a claim that it came earlier. He is also positive about embedding conventional Web solutions into agents, but questions tying that approach specifically to MCP. The post offers no further developments; it ends at the author's comparison and stated preference between the two routes.