Summary of "The New Application Layer - Malte Ubl, CTO Vercel"

Summary — “The New Application Layer” (Malte Ubl, CTO Vercel)

Agents are a new kind of software that both build and use software, creating a new application layer.
This makes many previously uneconomic automations viable, increasing the total amount of software produced and the demand for engineers.
Two possible futures:
- Model-lab dominated: AI remains expensive and value accrues to model providers.
- Commoditized models: models become cheap and innovation happens at the application/agent layer.
Malte argues the commoditized-models future is more likely and desirable.

Chat SDK: connector to hook agents into chat platforms (Slack, Telegram, WhatsApp, etc.).
just-bash: a Bash interpreter implemented in TypeScript that provides a fast, sandbox-like environment for agents.
Vercel’s in-house agents on vercel.com:
- Routing “contact sales” submissions by inspecting LinkedIn/company size and routing appropriately.
- Triaging abuse reports.
In-house support agent with ~90% deflection, handling routine queries in real time and routing only complex cases to humans, which improved job satisfaction.

Support automation: 24/7 agents handle predictable support work, offloading routine tasks to increase availability and reduce toil.
Compressed research: agents gather context and data (web/LinkedIn checks, docs) and leave final decisions to humans — low risk, high ROI because decision ownership remains with humans.
Information surfacing: extract and synthesize information already present across issue trackers, recordings, docs, Slack, etc., to make it practically useful.
Eliminate boring work: identify tasks people hate about their jobs and automate them to improve morale and productivity.

Agents are becoming major software consumers — Vercel observed >60% of page views on vercel.com coming from AI agents in a recent week.
Expect usage patterns to shift from human/GUI interactions toward APIs and CLIs.
Design for agents-as-users: expose clear APIs/CLIs and automation-first workflows rather than only UI-driven features.
Sandboxing and fast startup environments are essential; many teams ship sandboxes so agents can safely execute code or interact with systems.
Security is a major emerging challenge, comparable to earlier eras of widespread insecurity; anticipate paradigm shifts in architecture and defenses.

Separation of harness vs. execution: many agent harnesses are architecturally wrong because they mix where the harness runs with where generated code runs.
Separate orchestration/harness from execution environments — Malte notes Anthropic has adopted this separation in a recent product.
Be open to future paradigm shifts — current agent patterns are still early innings.

Cheaper software creation via agents will lead companies to “make” rather than “buy” (the “SaaS copocalypse” idea), increasing engineering work rather than reducing it.
Europe is well positioned for application-layer AI engineering innovation (examples: Vercel’s AI SDK team in Berlin, Pi — an Austrian coding agent, OpenClaw).
Europe may not lead model labs but can lead innovation at the application/agent layer.
Numeric and subtitle-derived claims from the talk may be inaccurate.

Focus on low-hanging agent use cases that speed up existing processes (research, routing, surfacing information).
Prioritize APIs/CLIs and sandboxed, separable execution for agent workloads.
Use the “what do you hate about your job?” heuristic to discover impactful automations.
Prepare for security and architecture changes; avoid coupling harness and execution.

Main speaker: Malte Ubl — CTO, Vercel (keynote presenter).
Projects/companies mentioned: Vercel (agents on vercel.com, AI SDK), Anthropic (agent product with separation), Pi (Austrian coding agent), OpenClaw.
Other people referenced in the talk (transcribed): Last/Lars Gammel (leads Vercel’s AI SDK — name/numbers may be mis-transcribed), Mario (speaker about Pi), Peter (speaker for OpenClaw).