Running AI Engineer with AI — swyx
TL;DR
Swyx says the real payoff isn’t more AI-written code — it’s more productive humans — after wiring Devin into AI Engineer’s workflows, his non-technical team started shipping animations, polish, and side ideas they previously wouldn’t even attempt.
AI Engineer itself is a case study in the ‘tiny team’ thesis — Swyx says the conference business has 9 full-time employees and does more than $9 million in revenue, which fits his definition of a team with more millions in revenue than employees.
The first breakthrough was Figma-to-production in ‘very short order’ — a designer expected a 1-4 week handoff cycle, but once Devin was connected to Figma via Coworker, it produced a pixel-perfect website that became the live ai.engineer site.
Agents changed collaboration patterns across time zones without a playbook — Swyx would kick off work, go to bed, and his Indonesia-based designer would wake up and keep prompting Devin with red-line annotations, using it like they’d communicate with any human teammate.
He argues agents kill ‘yak shaving,’ not just coding toil — the underrated benefit is parallelism and dependency cleanup, especially boring setup work like dependency trees and Python environment issues that usually block momentum.
Swyx is pushing beyond coding into ‘agents for everything else’ — he uses agents to manage conference schedules for 130 speakers, ETL from external vendor systems, research speakers from Apple Notes into Notion, and even source a literal lobster prop in London.
The Breakdown
The annual Swyx thesis, now pointed at his own company
Swyx opens by framing this as the latest chapter in a running AI Engineer keynote arc: first AI productivity, then falling model costs at roughly 100x every 12-18 months, then “tiny teams.” He reminds the crowd that AI Engineer itself is one: 9 full-time people, more than $9 million in business, and a very intentional attempt to prove small teams can punch way above their weight.
From a totally non-AI stack to agents in the company Slack
He says the irony is that the conference was originally run on a very normal stack — Figma, React, Supabase, Tldraw, Google Sheets, Sessionize — with nothing especially AI-native about it. The shift started when he joined Cognition, began seriously using coding agents because they were free, and casually dropped Devin into the company Slack to show teammates how it could help on the website.
The Figma-to-site moment that made it real
The big turning point: a contract designer showed him a Figma page that normally would have meant a 1-, 2-, or 4-week wait to make real. Swyx hooked Devin up to Figma — with Coworker handling the annoying setup — and suddenly they had a pixel-perfect working website, which he says is basically the site live today at ai.engineer. His takeaway is memorable: agents don’t just automate coding, they eliminate the “yak shaves,” the endless dependency and setup chores that quietly eat whole days.
When the team started using agents without being taught
After that first win, usage exploded — he flashes a Slack thread with 207 replies as proof. The part that clearly surprised him most was social, not technical: he’d start a task, go to bed, and his designer in Indonesia would wake up and continue the work by prompting Devin with red-line annotations, with no manual and no training doc, just the instinct to talk to it like a coworker.
Why agents made people do more work — happily
He tells a fun example: he saw a viral tweet about a design aesthetic, tossed it into Devin as a throwaway experiment, and it spiraled into a 127-reply thread and a real feature. The key insight for him is that the team now works on easter eggs, animations, and polish simply because it’s fun; the feedback loop is short enough that ideas don’t die waiting on a developer. That’s why he stops talking about lines of code and starts talking about “getting more productivity out of my humans.”
Running a 1,000-person conference like it’s a data pipeline
From there he broadens the scope: the conference is really a giant data-management problem involving 130 speakers, sponsors, attendees, and constant changes. He says the unlock was throwing away the CMS, using code as the source of truth, and letting Devin manage the schedule so a speaker email like Marta’s can be handled with “Devin, handle it for me.” That workflow, he argues, is what lets a team of nine run a 1,000-person event now and potentially 6,000 in San Francisco without growing headcount.
Agents for ETL, procurement, and replacing bits of SaaS
He runs through more “everything else” examples: syncing messy external vendor data into a single source of truth, turning tiny Apple Notes into a polished Notion-style speaker research doc with Glean/Tana-style tools, and even sourcing a lobster in London after seeing a viral Wall Street claw stunt. That lobster at the conference, he says, was effectively bought by Devin — a joke with a serious point that coding agents are breaking containment and becoming general-purpose knowledge workers.
The bigger trend: build for agents, not dashboards
He closes by tying his workflow changes to a broader industry shift. Citing Malte’s keynote that 60% of Vercel’s user base is now bots or agents, and talks from Edo and Liad on MCP apps, he argues dashboards matter less while APIs, CLIs, MCPs, and “agent experience” matter more. His final line is the whole thesis in one punchy phrase: agents for everything else are coming, so wake up and use them at work.