Agents for Everything Else — swyx

How much of running an AI conference can the organizers hand to AI? At AI Engineer’s first London event, Shawn “swyx” Wang approaches that question through the work of his own team. The opportunity extends beyond writing software: it includes the coordination, setup, and routine knowledge work that determine how much a small organization can accomplish.

His successive AI Engineer keynotes trace the path toward that question. The first, three years earlier, concerned productivity gains from using AI. The next encouraged greater use as costs fell. Swyx describes a roughly 100-fold decline in AI costs every 12–18 months as his working heuristic, rather than a guaranteed rate for every workload.

The third theme was tiny teams: teams whose revenue, measured in millions, exceeds their employee count. The accompanying Tiny Teams Playbook, developed around a World’s Fair track, uses the more specific measure of annual recurring revenue. The attraction is organizational, not just entrepreneurial. A large company can contain tiny teams; the idea does not require finding a lone founder who builds a billion-dollar business.

Swyx reports that AI Engineer, led by Ben Li and himself, has nine full-time people and runs a business exceeding $9 million. He describes that figure as business size, not ARR. The workflow changes since the organization began two years earlier become his practical example of adopting AI wholeheartedly—complete with physical “AGI pills” given to attendees. His joke is that everyone can now prescribe one to a coworker who remains unconvinced.

The original conference stack was stable and largely non-AI: Figma, React, Supabase, Tito, Google Sheets, and Sessionize. Joining Cognition gave swyx access to coding agents at work, and free access helped prompt serious experimentation. He added Devin to the company’s Slack and began showing colleagues how to request changes to the website.

Then the designer brought him a Figma page. The expected implementation cycle was measured in weeks—perhaps one, two, or four. Swyx brought Devin into the conversation, but first faced a prerequisite: connecting it to Figma. He delegated that setup to Claude Cowork. Before the coding agent could do the visible work, another agent helped remove the work blocking it.

An agent’s value includes the prerequisites it removes. Dependency installation and broken Python environments are concrete examples of this “yak shaving”: each intended action exposes another task that must happen first. Counting only autonomous task completion misses both the depth of that prerequisite chain and the benefit of handling work in parallel.

Once the connection was in place, swyx reports that Devin produced a functioning website matching the Figma design pixel for pixel. He does not give an elapsed implementation time. Seeing the result himself changed his confidence in the tool, and the designer was happy with it. Swyx identifies the result as the conference website live at the time of the talk.

The initial success turned into ongoing collaboration. Swyx points to 207 replies in one Slack thread as evidence of growing use. He would start work before going to bed; the designer, based in Indonesia, would wake up and continue with Devin. The designer began sending red-line annotations, communicating changes visually in the same way he might with a human developer. Swyx compares this with Steve Ruiz’s tldraw examples, but says he had neither taught the designer the technique nor supplied an instruction manual. For a mostly nontechnical team, that independent adoption mattered.

The team also began making things it would normally leave undone. Swyx added a website Easter egg that appears when a visitor on an ultrawide display moves the pointer over the highlights. Then he fed Devin a viral tweet whose design aesthetic he liked, simply to see what it would produce. That experiment grew into a thread with 127 replies. He had initially wanted to avoid spending the designer’s time on a throwaway idea, but the designer joined anyway.

The reason was enjoyment. The wait between an idea and a visible result had largely disappeared: the designer no longer had to block on swyx or a contract developer to try something. Animations, polish, and playful experiments became work the team could initiate directly. The productivity gain showed up in human initiative, not just generated code. This is the shift swyx calls agents for everything else.

Figma-to-website and tweet-to-website were useful successes in their own right. Together, they also changed the next question from whether the agent could help to where else the team could use it.

A conference is also a large data-management problem. Swyx describes coordinating 130 speakers, a couple dozen sponsors, and attendees with varied needs. Much of that work resembles maintaining a content management system: collect information, reconcile changes, and keep the published schedule current. The team had experimented with Sanity, but swyx found a more useful arrangement by supplying spreadsheets to Devin and moving the source of truth into code.

With the schedule represented in code, a coding agent could maintain it. Swyx says Devin managed the entire schedule, while emphasizing that another coding agent could fill the same role. His example is a speaker-change email from Marta. Instead of translating the email into a manually edited CMS entry, he could forward it—or paste a screenshot—and ask Devin to handle the change. The incoming message supplied the task context; the agent worked against the schedule’s existing source of truth.

Swyx describes a nine-person team managing a conference of roughly 1,000 people. That is his account of operational scale, not an audited attendance count. He then projects handling 6,000 people in San Francisco that summer without increasing the team size. His confidence depends on sufficient onboarding and established workflows: the throughput comes after the agent has been integrated into how the organization works.

The same approach extends to ETL, the extraction, transformation, and loading of data between systems. An external vendor holds information missing from the central source of truth. The task is to obtain the vendor’s API key, synchronize its data, and reconcile the information into that shared source. This is routine integration work, but it is also part of the conference operation that an agent can take on.

The next example leaves software maintenance entirely. A viral post showing a claw beside the Wall Street bull inspired swyx to put a similar prop in front of the conference. He asked Devin: “Where can I get a lobster in London?” The agent returned phone numbers, email addresses, and websites. Swyx clicked through the options, considered them, and requested more research. He credits that process with sourcing the lobster used at the conference; the described workflow includes his evaluation of the results, rather than an unattended purchase.

For this kind of task, his requirements are simple: web access and a sufficiently capable model. Whether the assistant is called OpenClaw, NanoClaw, or something else matters less than whether it can do knowledge work he would otherwise assign to an executive assistant or junior employee. A coding agent becomes an on-demand worker for research and coordination as well as implementation.

Swyx broadens the examples beyond Devin, disclosing that he now advises Cognition rather than working there in the capacity described earlier. He has begun exploring Town as coding agents expand into general knowledge work. He connects this direction with purpose-built knowledge tools, Andrej Karpathy’s discussions of wikis, and what he describes as OpenClaw’s adoption of that approach. He expects this expansion to be one of the leading trends of 2026.

For World’s Fair planning, he starts with Apple Notes containing prospective speakers and tracks. He supplies those notes to Town and gets a formatted Notion document containing research on the speakers he is considering approaching and curating. The names on the displayed notes are intentionally small; the useful part of the example is the transformation from a personal planning list into researched material that supports the next decision.

Success with individual tasks can lead to a larger ambition: replacing an entire SaaS product with something built internally. Swyx shows himself debating an unnamed replacement with employees. He jokingly calls his stronger enthusiasm “psychosis,” but takes the disagreement seriously. A manager cannot dismiss the objections of the people who will have to operate the result.

Those concerns are valid because employees bear the consequences when the replacement fails—and swyx explicitly acknowledges that the team gets things wrong. His method is to identify the three most important concerns and systematically reduce them. At the time of the talk, that evaluation is still underway. The example is an approach to earning confidence in a replacement, not a completed story of eliminating a SaaS dependency.

These internal changes also suggest a product-design problem: the entity using a service may increasingly be an agent. Swyx attributes to Malte’s opening keynote a claim that 60% of Vercel’s user base consists of bots or agents. The denominator and measurement period are not established here, so that figure should not be treated as a measured share of customers. It motivates swyx’s broader emphasis on APIs, CLIs, and MCP interfaces as ways for agents to use software.

There is a related change in where interfaces appear. Swyx points to MCP Apps speakers Ido and Liad, also referencing their ETN appearance, as advocates for shipping UI into someone else’s application. MCP Apps supplies a concrete setting for that idea: users can encounter a service’s interface within an MCP client. A product’s own dashboard no longer has to be the sole destination for using it.

Agent experience means designing for this changing primary user. Swyx’s own Figma workflow makes the shift tangible: he cares less about navigating the dashboard himself when he can give the task to Claude Cowork and hope it completes the work. His closing invitation is to bring that experimentation into everyday work, including the tasks surrounding software development. The AGI-pill joke returns as an invitation to bring hesitant colleagues along.