AI
AI Engineering as the Successor Discipline
The question Ubl gets most often: is there still a place for engineers? His answer is structural. Agents make software economically viable that traditional code never could. The Venn diagram of “software that should exist” was always larger than the Venn diagram of “software actually built.” Agents close that gap. As the cost to build falls, companies shift from buying SaaS to building themselves. More software made means more demand for engineers, not less. Ubl frames it as a speed-run of an economics experiment, and says there are no signs of an S-curve yet.
Four Agent Archetypes That Work Today
Ubl identifies four agent archetypes that are already working. First: 24/7 support agents (agents don’t sleep). Second, compressed research, an agent handles the research phase of any business-event-to-decision loop. Vercel’s sales intake agent filters 75% of contact-sales messages as support tickets; the other 25% get LinkedIn lookups and company sizing before a human touches them. Third: surfacing existing information from Slack and meeting recordings into issue trackers. Fourth: eliminating work people hate. Vercel’s in-house support agent achieves a 90% deflection rate, and job satisfaction on the support team rose because staff only handles cases that need real judgment.
Software Is Now Built For Agent Consumers
Over 60% of page views on vercel.com in the last seven days came from AI agents.
“humans are actually now in the minority.” — Malte Ubl
That shift changes what “shipping a feature” means. Ubl now asks every team: what’s the CLI? UI is cheap; agent-accessible APIs matter. Security is the third pressure. He compares today to 1999, when everything could be hacked. Most agent harnesses combine the harness runtime with the code-execution environment, which he calls fundamentally the wrong architecture. Anthropic’s new agent product separates the two, which he flagged as the right call.
Model Commoditization and the Application Layer
Two futures. In the first, a model lab wins, AI stays expensive, and engineers become “forward deployed” for whoever controls the stack. Ubl thinks that’s unlikely.
“The model companies are commoditizing.” — Malte Ubl
Google’s cheap infrastructure guarantees a price floor that stays low. In the commoditized future, the application layer captures the business value. Vercel’s AI SDK crosses 10 million weekly and is run out of Berlin. Pi, a coding agent built in Austria, and Open Claw are both European. Ubl’s argument: you don’t need to build models to lead AI engineering.
Notable Quotes
agents are a new kind of software. Malte Ubl · ▶ 4:05
humans are actually now in the minority. Malte Ubl · ▶ 13:05
The model companies are commoditizing. Malte Ubl · ▶ 17:45
Key Takeaways
- Cheaper software production expands total demand, so engineering work grows rather than disappears.
- Compressed research agents cut a 30-minute human task to 5 minutes without changing the decision process.
- As models commoditize, AI engineers owning the application layer capture the business value.