Swyx on Agent Loops and the Limits of Token Maxing

At the AI Engineer World's Fair—the conference swyx co-founded and now runs as something close to a barometer for where the developer community actually is—they ran an experiment. They called it the Token Billionaire program: spend a billion tokens a week, get a gold card, access to a special lounge. They expected fifty people out of seven thousand attendees to qualify.

They got three hundred.

That number landed differently than anticipated. Not because token consumption is inherently meaningful, but because of what it implies about the pace at which serious AI builders have normalized scale that was almost unimaginable eighteen months ago. Swyx told David Ondrej in a recent conversation that Cursor, the AI coding IDE, reportedly pushes eight to ten billion tokens per day. "Eight is very high," he noted, with the measured skepticism of someone who's talked to enough teams to know when a number is real.

This is the context in which swyx's central argument lands: that the prompt-writing era is giving way to something structurally different. Not just better prompts, but loops—automated systems where prompts are generated by other prompts, where the human is no longer the one sitting at the keyboard asking questions but instead the architect of systems that ask questions on their own.

Loops all the way down

The framing comes from Boris Cherny's work, which swyx cited approvingly: you should not be writing prompts, you should be writing a loop that creates prompts. It's an abstraction jump that sounds cleaner in theory than it is to implement in practice, but the core intuition is real. The more capable models get, the more your value-add shifts from "knowing exactly what to ask" to "knowing what outcome to optimize for and building the machinery that iterates toward it."

Swyx laid out what he thinks a well-designed agent loop actually requires: a specification (what are you looking for?), a verification mechanism (how do you know when you're done?), and—the piece most people skip—a set of constraints (what should the agent not do?). That third component is where the real institutional knowledge lives. If your model keeps generating 10,000-line files that no one can maintain, you don't fix that by hoping the next model will know better. You encode the constraint and move on.

The practical entry points he suggested were concrete: an "interview me" loop that flips the dynamic, getting the model to ask you questions rather than the other way around. ("We're actually very bad at prompting. We're actually very bad at expressing what we want.") A competitive research loop that runs weekly, generates ideas, prototypes them, and throws most of them away. Closing the SDLC loop by feeding production error logs directly to your agent so it can propose fixes before you've even noticed the bug.

The north star version of all this is what he called the "self-healing app"—a codebase that detects its own problems and patches them without waiting for a human to notice. Whether that's a useful design goal or an invitation to runaway agent behavior is, genuinely, an open question.

What you can't hand off

The part of this conversation worth sitting with longer is where swyx draws the line.

He's emphatic that token volume without user contact is just a liability. "Yes, you're spending a billion tokens a day, but you goddamn better be actually having real humans be using your app, because it's super super easy to code a lot of slop that makes your app completely unusable." He's not being moralistic here—he's being operational. Models don't carry the accumulated intuition of someone who's used ten similar apps and knows, without being able to explain why, that the button placement is wrong. That intuition isn't in the training data in a usable form. It's in your users.

The deeper version of this is the data structures argument. His advice to anyone building seriously with AI: never lose track of your data layer. Not the UI, not the agent orchestration, not the prompt templates—the schema. What gets logged, what gets reported, what the workflows actually operate on. If the data structure is wrong, no amount of agentic cleverness fixes it. You just burn two months and throw everything away.

What he's describing is a division of cognitive labor that's still being worked out in real time. Agents are good at generating, iterating, researching, and executing defined tasks at scale. Humans are still necessary for taste, for the "I know it when I see it" judgment that determines whether what got built is what anyone actually needs. The token-maxing dynamic creates its own pressure here—loops are, by design, incentivized to consume tokens, which means they're incentivized to do things, which means without human oversight they'll keep doing things regardless of whether those things are useful.

The model landscape, briefly

Swyx's take on the AI model market is more nuanced than the "big labs win everything" narrative. His frame: frontier labs (OpenAI, Anthropic, Google) will always own the frontier because that's where the GPU clusters and research talent accumulate. But most work isn't frontier work. Most work is boring, routine, domain-specific. That's the space for what he calls "agent labs"—companies like Cognition that train models specifically for enterprise coding rather than general intelligence—and the space is large enough for multiple winners.

He made one observation that's worth flagging: the distinction between fine-tuning and post-training is collapsing in a meaningful way. When Mistral publicly disclosed that their post-training compute was four times their pre-training compute, the idea that post-training is a "thin layer" on top of a foundation model stopped making sense. The foundation is the post-training, increasingly.

On inference speed, he's bullish in a way that goes beyond the usual boosterism. The current baseline is 50-100 tokens per second from major labs. Cerebras-powered systems run at roughly 1,000. He's seen demos at 10,000, and claims to have seen demos at hundreds of thousands of tokens per second. "When you have 100 times faster inference, the type of products that you make will change." That's not a claim that faster is always better—it's an observation that certain product architectures are only possible when latency drops below certain thresholds.

On taste, founders, and lying

One thread in this conversation that doesn't get talked about enough in the AI developer community: swyx's argument that taste is the actual scarce resource, and that its absence usually presents first as a willingness to lie.

His example was the common demo-day claim that a 4B parameter model beats frontier performance. "You see a lot of this," he said, "and some of it might be somewhat legit, but in general people know you're full of it." The founders who make those claims aren't just wrong—they've revealed something about what they actually optimize for. They're maximizing the impression they make in the moment at the cost of the credibility they'll need later.

The positive version of taste, in his framing, is being more invested in the problem than in your company's position in the problem. The founders who teach you something about how inference bottlenecks work, who share two of their hundred trade secrets because those two are relevant to the conversation—those are the ones he finds worth paying attention to. "The best people with taste share something because they know they have a hundred trade secrets. They're only sharing two."

There's something almost governance-like about this argument. The people who move fields forward are the ones who contribute to the shared knowledge base while still protecting what they need to protect. It's not altruism—it's long-term positioning. But the effect is a healthier ecosystem.

The tension swyx leaves unresolved—and probably can't resolve, because nobody can yet—is whether this standard survives contact with the funding environment. When you're burning runway fast and need to close a round, the incentive structure points directly toward the behavior he's criticizing. Mission-driven founders make great role models. They're also easier to have when you're not staring at six months of cash.

The loops are getting more capable. The judgment about what to put in them is still, stubbornly, a human problem.

Dev Kapoor covers open source and developer communities for Buzzrag.