Vercel CTO: Agents Are the New Application Layer
Why Agents Are a New Kind of Software, Not Just a Faster Way to Write It
This is Malte Ubl, CTO of Vercel, opening the first AI Engineer conference in Europe. His keynote is not about models — it’s about the application layer being built on top of them, and why the economics of software itself are shifting.
On agents as a new software category: “There was always all this stuff we wanted to automate, but not all of it was economically viable to do with traditional software. But it is with agents.” Malte frames agents as filling a previously uneconomic corner of the software market — the workflows full of edge cases and tacit business knowledge that nobody would ever hand-code in a traditional SaaS product.
On the SaaS “copocalypse”: “More and more companies, when they ask whether they should buy some software or make some software themselves, they’re answering that with the make side.” His take: SaaS will survive, but more companies building more custom software means more work for engineers, not less. The software market is proving surprisingly elastic — the cheaper it gets to make, the more of it gets made.
On the four agent archetypes already working today: Malte lays out the practical patterns he sees shipping — not speculative research, but deployed systems:
- Agent-as-a-service / customer support — the 24/7 version of a 9-to-5 role.
- Compressed research — an agent does the research phase of a business process, and a human still makes the final decision. Vercel’s “contact sales” form is routed by an agent that reads LinkedIn, checks company size, and hands a briefed packet to a human.
- Surface existing information — “Is your issue tracker up to date? Probably not. Could it be? Yes — the info exists in Slack, in a Granola recording, somewhere.” Agents stitch existing company knowledge into the place it’s needed.
- Eliminate boring work — Vercel’s own support agent has a 90% deflection rate, and the team’s job satisfaction went up because humans handle only the interesting cases.
On agents as users of software: “In the last 7 days, over 60% of page views on vercel.com were AI agents.” The first-order consequence is that UIs are getting cheap while CLIs and APIs are getting expensive. The second-order consequence is that infrastructure has to run software that nobody on the team actually wrote.
On where value accrues: “Model companies are commoditizing. In that world, we the AI engineers are the powerful ones. Our agents are the ones that actually create the business value.” Malte is explicit that Europe won’t win the model war — but it doesn’t need to, because Vercel’s AI SDK, Poe, and OpenCode are all European-led and sitting in the layer where the money now lives.
Four Agent Archetypes Every Business Can Ship Today
- Agent-as-a-service - The support role that runs 24/7 (CRMs, Decagon pattern)
- Compressed research - Agent researches, human decides; 30-min job becomes 5-min, 100K times a year
- Surface existing info - Stitch info from Slack, recordings, docs into issue trackers and status updates
- Eliminate boring work - 90% deflection agent; support team’s job satisfaction exploded
- Ask “what do you hate most about your job?” - Malte’s magic question for finding agent opportunities
- 60% of web traffic is already agents - UIs are cheap; CLIs and APIs are the real interface now
What This Means for Companies Building Agent-Powered Workflows
The application layer is where the next decade of software gets built and where the economic value accrues. Companies don’t need to pick the “winning model” — they need to pick the workflows worth automating. Malte’s framework of compressed research and boring-work elimination is exactly the pattern that turns AI from a curiosity into a line item: same process, same risk profile, an order of magnitude less human time. For anyone hiring AI employees rather than buying another SaaS seat, these archetypes are the playbook.