Agents for Everything Else: Coding Tools Break Into Knowledge Work

swyx's closing keynote at AI Engineer London (May 1) argues that "agents for everything else" — coding-agent harnesses applied to non-coding knowledge work — is the most under-discussed trend of 2026. His proof case: the AI Engineer conference itself, run by nine full-time staff, generating $9M in revenue, and scaling to 6,000 San Francisco attendees this summer without adding a single hire.

What the Source Actually Says

In the keynote transcript, swyx details how Devin (Cognition Labs) now runs the AI Engineer stack: Figma-to-website translation — pixel-perfect and signed off by an Indonesian designer who learned to prompt it via TLDraw red-line annotations, with no instruction manual — plus schedule-as-code replacing the Sanity CMS, ETL between vendor systems, and physical logistics. Devin researched 3D claw rental options in London, surfaced phone numbers and emails, and the conference prop was sourced on the back of that output.

The structural argument is more important than the anecdotes: agents eliminate yak-shaving — the dependency-installation tax that compounds at every "first I have to…" step — and unlock parallelism that conventional productivity models miss entirely. "I am getting more work out of my employees because they enjoy doing it," swyx said. The feedback loop from idea to execution no longer requires him as a bottleneck, producing animations and polish the team would never previously have prioritised.

Karpathy's independent Sequoia Ascent 2026 fireside (surfaced April 30 on X) provides structural framing: the "agent-native economy" decomposes products into sensors, actuators, and logic across Software 1.0/2.0/3.0 paradigms, with "agentic engineering" emerging as a distinct discipline and hiring category. His summary: "vibe coding raised the floor; agentic engineering raises the ceiling." At the same AI Engineer event, Vercel CTO Malte reported that 60% of Vercel's dashboard traffic is now agents, not humans — CLIs, APIs, and MCPs are becoming the primary product interface.

Strategic Take

Teams should audit their non-technical knowledge-work stack: every SaaS tool a non-developer employee uses daily is a candidate for agent replacement. Capability is no longer the blocker — workflow habit is. swyx's practical method: identify the top three employee concerns and reduce them systematically, rather than steamrolling resistance from the people who will handle the downstream failures.