The Last Interface
Screens were a fifty-year detour. As models learn to negotiate the world on our behalf, the question is no longer how we talk to machines — it is what remains of the world they hand back to us.
The graphical interface is dying of success. For fifty years the screen was where humans met computation: we learned its menus, memorized its gestures, arranged our work around its windows. That arrangement assumed one thing — that the human had to do the operating. In 2026, that assumption is failing in public.
Look at what shipped in July alone. Anthropic's Opus 5 handles a one-million-token context and operates a computer well enough to complete most of a benchmark built from real desktop workflows [1]. OpenAI, Google, xAI, DeepSeek and Alibaba all shipped frontier updates inside six weeks [3]. The capability race is usually described as models getting smarter. The quieter story is models getting hands.
When software can read any screen, click any button and file any form, the screen stops being the product. What the user touches becomes a brief — an intent, stated once, in words — and the interface work happens somewhere they never see. The new agent products make this explicit: they live in Telegram, Slack, email and the command line, not in a destination app. The interface dissolves into whatever surface you already had open.
The web, renegotiated
This breaks a bargain the web has run on for thirty years. Sites earned attention with design, and monetized the visit. An agent does not see design. It sees structure, prices, claims. It does not linger, cannot be retargeted, and does not click ads. A meaningful share of page views is already synthetic; when the buyer of information is a machine, publishing economics reorganize around being machine-legible — clean data, verifiable claims, primary sources — rather than sticky.
Regulators have noticed before most product teams. The EU's transparency obligations, enforceable since August 2 with fines up to €15 million or 3% of global turnover, require that synthetic content be marked and that people be told when they are talking to a machine [2]. The disclosure rule is a tell: lawmakers now treat 'is this a human?' as a question ordinary consumers can no longer answer unaided.
What remains
Interfaces do not vanish; they concentrate. Someone still writes the brief, sets the constraints, and reads the result. The scarce skill shifts from operating software to specifying outcomes and auditing what came back — the same shift that made management a profession when work industrialized. The last interface is not a screen. It is the sentence you give the machine, and the judgment you apply to what it returns.