Agents for Everything Else — swyx

Swyx, co-founder of AI Engineer conferences, delivers a closing keynote reflecting on how his nine-person team uses AI agents to run a conference business generating over $9 million in annual revenue. The talk is framed around his evolving keynote arc: year one covered productivity gains from AI adoption, year two emphasized the AI cost curve dropping roughly 100x every 12–18 months, year three introduced the 'tiny teams' concept (teams with more millions in revenue than number of employees), and this year's thesis is 'agents for everything else' — applying agents to operational and knowledge-work tasks beyond pure coding.

Swyx describes how joining Cognition and gaining free access to Devin triggered a fundamental workflow shift. The first major use case was Figma-to-website: a designer in Indonesia was preparing a Figma page and expecting weeks of developer turnaround. Instead, Swyx connected Devin to Figma (using a tool called Coworker to handle the integration setup), and the team shipped a pixel-perfect, fully functioning website in very short order. That site is the live ai.engineer website today.

A key observation emerged from watching a non-technical team member interact with Devin autonomously: the designer began annotating Figma frames with red-line markup, mirroring how Steve Ruiz uses TLDraw annotations, without any instruction manual. The designer started working on features that would previously have been deprioritized — including Easter eggs on the site and viral design aesthetics pulled from tweets — simply because the feedback cycle was instant and the work felt enjoyable. Swyx argues this is an underappreciated mechanism: agents increase output from existing human employees not just by automating tasks but by making previously-blocked or low-priority creative work accessible and fun.

The second major operational use case is conference schedule management. Swyx describes the entire AI Engineer schedule as a giant data management problem: 130 speakers, dozens of sponsors, all attendees. His solution was to abandon a CMS (specifically noting dissatisfaction with Sanity) and commit the schedule data to code, treating that code as the source of truth and letting a coding agent manage it. Speaker changes — such as a last-minute request from a speaker named Marta — are handled by forwarding an email to Devin with no further instructions. This approach is what allows nine people to manage a 1,000-person London event and plan for 6,000 attendees at the San Francisco World's Fair.

Swyx also describes using agents for ETL (syncing data from an external vendor system to a central source of truth), for research (throwing Apple Notes with speaker names into Town/Notion assistant to generate formatted research documents), and even for procurement — asking Devin to research where to buy a lobster in London, which resulted in the actual lobster sculpture that appeared at the conference, purchased based on Devin's research output of phone numbers, email addresses, and websites.

The talk highlights a broader macro trend: Malte Ubl's opening keynote noted that 60% of Vercel's user base is now bots and agents, not humans, which means APIs, CLIs, and MCPs matter more than dashboards. Swyx echoes this by citing Edo and Liad's talk on MCP apps and the shift toward 'agent experience' (AX) as the primary design target, with custom UIs becoming less important than shipping UI into other agents' surfaces.

Swyx closes with a challenge: identify the top three concerns your team has about replacing SaaS tools with agent-built alternatives, then systematically reduce those concerns. He acknowledges managing employees who are less 'in psychosis' about AI replacement requires respecting their valid operational concerns. His final call to action is direct: use agents for everything else, bring it into your workplace, and if colleagues are insufficiently convinced, prescribe them an 'AGI pill.'