The New Application Layer - Malte Ubl, CTO Vercel

{ "tldr": [ "AI engineering is the successor to web development — Malte Ubl argues that after the disruption of how we build software and what we build, AI engineering is becoming the mainstream discipline that will define the next decade.", "Agents make previously uneconomic software worth building — his core thesis is that there has always been a huge set of software that “should exist” but was too expensive to hard-code, and agents now make that long tail viable.", "The best near-term agents are boring in a good way — instead of chasing fully autonomous systems, Ubl points to practical wins like 24/7 service agents, “compressed research” workflows, and internal knowledge surfacing that save companies millions without changing core processes.", "Vercel is already seeing agents become the users of software — over the last 7 days, more than 60% of page views on vercel.com came from AI agents, and product usage is shifting from dashboards to APIs and CLIs that agents can automate.", "Security and architecture need a reset for agent-native systems — Ubl says many popular agent harnesses are built wrong because they mix the harness with the generated code runtime, and warns the industry is heading into a “1999”-style security wake-up call.", "Europe may not win on foundation models, but it can lead the application layer — he highlights Vercel’s AI SDK, Austria’s Pii, and OpenClaw as signs that Europe is already strong in AI engineering, especially if models become cheap commodities rather than monopoly products.", ], "breakdown": "### A keynote, not a technical talk — but still very Malte\n\nUbl opens by joking that he usually gives technical talks, but since he’s first on stage he felt obligated to do a “proper keynote.” Even then he can’t resist name-dropping his own “vibe coding stack,” including Chat SDK for wiring agents into Slack, Telegram, or WhatsApp, and “just bash,” a TypeScript bash interpreter with near-instant startup because, as he puts it, agents love bash.\n\n### Why AI engineering feels like the next big community\n\nHe reflects on shutting down JSConf EU in summer 2019 with “impeccable timing” right before COVID would have killed conferences anyway, and says it’s exciting to finally see Europe’s tech community reconvene. But he’s clear this wasn’t going to happen around plain old web development in 2026 — AI engineering, in his view, is the legitimate successor and the discipline now pulling the whole industry forward.\n\n### Will engineers still matter? His answer is yes\n\nUbl says he’s constantly asked whether there’s still a place for engineers, especially junior ones, and answers with a TikTok analogy: he’d be terrible at it because he didn’t grow up with it. In the same way, he thinks the next generation will be naturally better at AI-native engineering — but older engineers will be fine too, because agents expand the total amount of software the world can economically justify building.\n\n### The real thesis: agents fill in the software that never got written\n\nHe sketches a mental Venn diagram of “all software that should exist” and argues that traditional engineering left a huge chunk undone because encoding every business rule was too costly. Agents change that math, which is why he sees the market as an economics experiment in elasticity: the cheaper software gets to make, the more of it companies will make, not less.\n\n### The practical agent patterns already working today\n\nRather than hyping moonshot autonomy, Ubl focuses on what’s shipping now: support-style agents that can work 24/7, “compressed research” agents that do the investigation before a human decision, and systems that surface information already scattered around the company. His Vercel example is especially concrete: contact-sales submissions first hit an agent, which reroutes about 75% of them to support and enriches the real sales leads by checking LinkedIn, company size, and routing before a human ever steps in.\n\n### Ask people what they hate about their jobs\n\nHis favorite prompt for finding agent opportunities is brutally simple: ask employees what they hate most. At Vercel, an in-house support agent now deflects 90% of requests, and the support team’s job satisfaction jumped because they stopped dealing with repetitive chores like failed credit cards and got to focus on the genuinely interesting cases.\n\n### Software is now being used by agents, not just humans\n\nUbl says developer tools are already ahead of the curve here: in the last 7 days, over 60% of page views on vercel.com came from AI agents. That’s changing product design itself — when someone proposes a feature with a UI mockup, his first question is now “what’s the CLI?” because dashboards are no longer the primary interface if an agent is the real user.\n\n### The infrastructure and power shift ahead\n\nHe warns that agent-native systems need new infra, new sandboxing patterns, and much better security, comparing the current moment to 1999 when “everything can be hacked.” He also thinks many agent harnesses are architecturally flawed because they run the harness and generated code together, then closes on a geopolitical note: Europe may not dominate model labs, but if models commoditize — with Google helping drive prices down — then the real leverage moves to the application layer, where AI engineers, not the labs, create the business value.", "oneLiner": "Malte Ubl’s case is that agents are creating a new application layer where the biggest opportunity isn’t building foundation models, but building the software, workflows, and infrastructure that make AI economically useful everywhere.", "tags": ["industry", "commentary", "product"] }