The talk is “The New Application Layer” by Malte Ubl, CTO of Vercel, delivered at AI Engineer (likely the Europe edition or World’s Fair around April 2026).
Ubl argues that AI engineering is the legitimate successor to web development and will be the mainstream discipline shaping the next decade. As foundational AI models commoditize, economic value shifts decisively to the application layer, where AI agents (both as builders and users of software) drive innovation, unlock new automation, and reshape infrastructure. He draws on Vercel’s experience and data to illustrate this transition.
Core Thesis: Economics of Software Elasticity
- AI dramatically lowers the cost of creating software, leading to far more software being built overall (an “experiment in economics of how elastic the software market is”).
- This doesn’t reduce demand for engineers; it increases it. Agents make previously uneconomical or complex/context-dependent software viable, filling gaps that rule-based systems couldn’t address.
- Enterprises increasingly build custom internal tools with AI rather than buying SaaS (“SaaS copocalypse” narrative), turning more organizations into software creators.
- Quote: “We are speed running what’s really an experiment in economics of how elastic the software market is: the cheaper it is to make software, the more software we are going to make, and the demand for software engineers is going up.”
Evidence from Vercel
- Over 60% of page views on vercel.com recently came from AI agents, meaning agents are now the majority “users” of the platform. This shifts priorities from human-centric UIs to robust APIs, CLIs, and agent-ready interfaces.
Pragmatic Enterprise Agent Archetypes
Ubl outlines four low-friction, immediately useful agent types for businesses (focusing on incremental integration rather than full replacement):
- 24/7 Workforce: Turns human roles into always-on operations.
- Compressed Research: Accelerates the research/prep phase of decisions while keeping human processes intact (e.g., qualifying sales leads from contact forms or triaging abuse reports).
- Information Surfacers: Extracts and makes usable buried company knowledge from Slack, meetings, recordings, etc. (e.g., auto-updating issue trackers or status reports).
- Toil Eliminators: Removes repetitive, hated tasks (e.g., an in-house support agent achieving ~90% deflection rate, boosting team satisfaction).
Designing Software for an Agent-Centric World
- UX/Infrastructure Shifts: Less developer attachment to generated code means infrastructure must “just run” reliably with minimal ops overhead. Focus moves to APIs/CLIs over graphical dashboards.
- Security & Architecture: Early-internet-like security challenges ahead. Most current agent harnesses have a critical flaw—executing generated code in the same environment as the harness (risking security/stability). Better: isolated sandboxes (e.g., as Anthropic’s agent product does).
- Agents as primary users demand “AI-first” APIs and designs optimized for automation.
Innovation Leadership and Future Scenarios
- Europe leads in AI engineering innovation (application layer independent of base models), citing examples like Vercel’s AI SDK (millions of requests/week), Pi (Austrian coding agent), and OpenClaw.
- Two futures:
- Model-centric (few big labs dominate; AI engineers become implementers).
- Commoditized models (Google et al. drive down costs; AI engineers and the application layer hold power and drive value/innovation).
- Ubl bets on the latter, positioning AI engineers (especially in Europe) to define the next decade.
Overall, the talk is optimistic and pragmatic: AI won’t replace engineers but will amplify them enormously, requiring new mindsets, architectures, and interfaces. The application layer—built and used by agents—is where the real innovation and value creation will happen. The ~18-20 minute keynote blends vision, data, concrete examples, and forward-looking advice for builders and enterprises.
For the full experience, watch the video directly, as it includes visuals like the agent harness flowchart.
