Sean Wang, known as swyx, who runs the Latent Space podcast and the AI Engineer conferences, said his company was unhappy paying about $40,000 for an events-software subscription that, in his words, nobody enjoys using. He decided to find out whether someone could build a replacement instead. He called the result Kill My SaaS: a bounty on what he described as "mid-tier SaaS that should not exist."
SaaS stands for software as a service: a program that a vendor runs online and rents to customers through a subscription. swyx said his events team agreed to switch away from its vendor. He did not describe the process as easy, and he did not claim that agents can now build whole products by themselves.
swyx discussed the bounty in an interview with Prakash, which appeared in the weekly highlights episode of The Cognitive Revolution published October 8, 2026. The interview also covered his view of what the computing stack lacks for AI agents. Nathan Labenz, the show's host, then pressed him on what this means for jobs and companies.
Salaries versus tokens
swyx explained the idea behind the bounty through a comparison. The subscription cost about "half a salary or one third of a salary" for software his company did not own. He said that money spent on tokens, the units of text that AI models read and write and that providers bill for, "buys me a heck of a lot of tokens."
The original event brief, presented by Latent Space, gives the terms. It offered a $10,000 prize for an open-source replacement for enterprise software that, according to the brief, his team was proposing to pay more than $40,000 a year for, without having used it and without being able to customize it. Entrants built their versions remotely, and swyx's team acted as the customer and judge. The brief was later updated because of the response. After more than 100 people signed up, the organizers stopped paying expenses in advance and instead let qualifying submissions claim up to $500. Those signups show interest in the bounty. They do not count finished entries.
"Evals are the problem"
The hard part, swyx said, came after the entries arrived. "We have so many submissions that we then have to eval them," he said. "And so evals are the problem." An eval is a structured test of whether something works. In this case the judges were not checking two or three requirements. They were testing the whole experience from three points of view: the event organizer, the attendee, and the sponsor or speaker.
He warned that a large prize attracts plenty of low-effort entries. Some were little more than an instruction like "Claude, make no mistakes, go do this," sent in without review. "The sort of verification load is very imbalanced because they spent zero thought on this thing," he said. An entrant could produce a submission almost for free, while the judges still had to spend real effort finding out whether it was any good. Even so, he called the bounty "very successful."
Winning over the spreadsheet crowd
The people swyx had to convince were his own events staff. He described them as "super old school" event professionals who "do everything in spreadsheets," work with unions and worry about where meter boards physically go at a venue. They distrusted anything new and "techie." At first their answer was "we'll never use this vibe coded thing." Vibe coding means building software mostly by describing what you want to an AI model and iterating on what it produces. The team said it would rather stick with the tried-and-tested platform used by big companies.
That changed when the team saw the quality of the submissions and compared them with the platform it was already using. swyx recalled the reaction: "yeah, okay, we're going to switch." He said the first hurdle was proof that a replacement was feasible at all.
The second benefit came after the decision. Through a partnership with Cognition, swyx said, he gave the whole team access to Devin, Cognition's AI coding agent, so they could change the code. Staff now request a change and usually have it within one to two hours, which he said they had never had before. He compared this with how SaaS vendors handle requests: a vendor says a feature is "on our Q3 roadmap," with no confidence it will arrive in that quarter. One change his team requested in the first quarter, he said, had only just landed.
"Quite cooked," with caveats
swyx's verdict was that "SaaS is quite cooked," but he added a condition: it applies "if you're mostly a CRUD app." CRUD stands for create, read, update and delete. It describes software that mainly stores records and lets people edit them. He said a lot of user-experience problems remain.
He also warned against reading too much into coding benchmarks. People see scores of around 90 on SWE-bench, a widely cited test for coding agents, and are impressed. "You're looking at the wrong thing, my guy," he said. The SWE-bench FAQ explains what the benchmark measures. Each task is a real GitHub issue in an existing code repository, and an agent passes if its patch fixes the issue according to tests. That is repair work on existing code, not designing and delivering a complete application that people enjoy using. His practical test was simpler. Anyone who has not tried to vibe code a complete version of a SaaS product they use, he said, does not "understand how models are still very bad at all this."
What the stack lacks for agents
Prakash had also asked which parts of the internet's infrastructure are poorly designed now that agents are becoming its main users. swyx began with computing hardware. Everyone knows about the shortages of GPUs and memory, he said, but a CPU shortage "is the most consistently reported thing among all the people that I interview right now." CPUs are the general-purpose processors that run the code around the AI models.
He said this calls for more "fluid compute," which he called Vercel's term. He meant short-lived, serverless computing that can run, pause and resume, because agents spend time waiting on model or network calls. He described the need as "long running states that are resumable." Vercel's own documentation on fluid compute describes something narrower. Several requests share one function instance, so the machine does useful work while one request waits for a database, network service or model response. The functions still have time limits, and Vercel points to a separate product, Workflows, for jobs that must pause and resume over much longer periods.
Next he named networking latency, meaning the delay in getting data from one place to another. This raises questions such as which region or access point an agent connects to and how much computing should happen in the cloud versus on local or edge devices. He said robotics companies are dealing with this now, and he called them "the first harbinger" of what others will face once they run enough inference, the work of having a trained model produce outputs.
His last item was personal: bandwidth, or "how many tokens per second can I get?" He meant that for a single call, for all his calls combined, and for his whole company.
The downside question
Labenz pushed back. He said he had trouble seeing how the industry would not pick off the easy targets and leave "a lot of companies go out of business." He pointed to layoffs at Meta as a possible "canary in the coal mine" and asked whether big tech would see mass layoffs. "How long can the party go on?"
swyx said the answer depends on how you frame it. If you treat software engineering as a fixed amount of work, he said, you reach Labenz's conclusion. But if the real competition, or total addressable market, is spreadsheets, the picture changes. Every spreadsheet anyone has built could become custom software with automations and a good interface. Beyond spreadsheets are phone calls, back-and-forth emails and physical meetings, which he said could gradually be turned into custom software, hardware and models. On that view, "we're just growing into the long tail." Many small, specialized uses that were never worth building software for would now get their own.
He used Salesforce as his example. It is big, he said, because people can customize it. But at some point customers will stop paying a "300k a year baseline subscription" and spend $30,000 on building their own customer-relationship software instead. "They're happier because they can now modify it to whatever they want."
Then he compared software to clothing. "We don't even have the same amount of customization in our software as we do our handbags," he said. People pick clothes by brand and style, not benchmarks. He said software engineering will only be "done" when there is a "Kate Spade versus Louis Vuitton versus Coach for software," where people choose on feel and identity rather than features or price.
swyx did not give a forecast for layoffs. He said the real limit is "the sum total of humanity's total token bandwidth," which he defined as "literally the amount of silicon we can produce." That supply will set prices, availability and rate limits, he said. The companies that matter most are focused on it, and "the rest of us are just fighting for scraps within the fixed pie that we already have."