Summary
David Ondrej interviews swyx (Shawn Wang), founder of the AI Engineer World’s Fair and an employee and adviser at Cognition (makers of Devin), for a candid practitioner’s perspective on how serious AI engineers are actually building with agents in 2026. The conversation draws heavily on firsthand data from the most recent AI Engineer conference, where roughly 7,000 attendees participated — and around 300 qualified as “token billionaires,” spending over a billion tokens per week, far exceeding organizers’ expectations of 50.
swyx argues that the most underrated trait in AI engineering right now is tolerance for “slop” — the willingness to run many parallel agent tasks, accept a high failure rate, and extract value from the useful fraction. He contrasts this with the default engineer instinct toward efficiency and cleanliness, suggesting that those who can stomach slop are still finding non-consensus advantages. He also observes that the role of the IDE is declining: tools like Devin, Claude Code, and Codex increasingly make the IDE just a file reviewer rather than the primary development environment.
The interview also covers product taste as a scarce resource in the age of cheap code generation, using the example of C++ creator Bjarne Stroustrup’s approach to language proposals — saying yes to individual features while saying no to the aggregate madness they create together. swyx frames this as the core challenge for AI product builders: knowing what not to build when building anything has never been easier. A thoughtful conversation for anyone building seriously with agents or trying to develop better product judgment in an AI-saturated environment.
📺 Source: David Ondrej · Published July 09, 2026
🏷️ Format: Interview
