Your agent will be your undoing

Get out of the token path.

This whole post in 33 seconds.

Kermit: Uh. That’s very effective.

—

If you want to make money, there may never be a more effective way of doing it than saying, “It is 2026, and I’m here to sell you artificial intelligence.”

Anthropic is selling so much artificial intelligence that its revenue has grown from $9 billion in January to $47 billion today, adding a full Salesforce of new business, in about five months. According to the Atlantic, “researchers at Goldman Sachs who conducted interviews with 40 software companies about their AI use in mid-April found that many were ‘overrunning their initial budgets’ for AI tools ‘by orders of magnitude.’” Uber burned through their annual Claude Code budget in four months. Nvidia now spends more on AI compute than on employee salaries.

And so, obviously: Nearly all companies are now AI companies. Photo-sharing apps are building $200-a-month vibe coding agents. Rocket ship companies are trying to impress investors not with rocket ships, but with AI. Shoe manufacturers are leasing data centers. Brian Chesky plans to start a new AI company. ChatGPT has a billion users.

It is 2026. Get in the token path. Sell artificial intelligence.

---

Gonzo: Yeah, it’s great when it works.

—

But how? How do you sell artificial intelligence? You could manufacture new chips, but that’s very expensive; you could train new AI models, but that’s very expensive and doesn’t always work. The better answer—or at least, the easier answer—is to sell tokens, by building an agent.

Because, first, it’s not so hard to build an agent. Agents, which are roughly synonymous with a “harness,”¹ are the applications wrapped around an LLM that fetch data, prompt the model, read its responses, re-prompt the model, and share stuff back to the agents’ users. This is all just software, and people have been wrapping software around databases and third-party services for decades. You need billions of dollars to build chips and models; you need a computer and a few days to build an agent.²

Second, despite their relative simplicity, agents can still be valuable products. According to a research paper from Meta, “changing the harness around a fixed large language model can produce a 6x performance gap on the same benchmark.” Though a small startup can’t outbuild Anthropic’s or OpenAI’s models, it can build a harness that improves their models.

Third, it’s also not that hard to come up with an idea for an agent. What do you do with a computer? Don’t do it anymore! Build an agent to do it! Agents are the first order effects of AI—we used to make documents, or manage emails, or analyze data, or look for dates. But what if agents did all of that instead?

And finally—and maybe most of all—agents can make money. Cursor built a coding agent a few years ago; it now makes more than $2 billion a year, and was recently bought by SpaceX for $60 billion.³ Cognition also built a coding agent; it now earns $500 million a year, and is worth $26 billion. Replit and Lovable built agents that make websites; together, the two companies make more than a billion dollars and are valued above $15 billion. Sierra sells an agent for customer support teams; it makes $150 million a year and is worth $15 billion. Harvey sells an agent for lawyers; it makes $200 million a year and is worth $11 billion.

And so, obviously: Nearly all software companies are now agent companies. Linear used to be an issue tracking application; now, issue tracking is dead; introducing Linear Agent. Introducing Scout, your always-on personal agent. The Figma design agent is here. dbt Wizard is the AI agent built for analytics engineers. Meet ZoomMate, your AI teammate. Meet Slackbot, your personal AI agent for work. Meet your Townie, your personal AI assistant.

It is 2026. It is the beginning of the harness era. Build an agent.

---

Beauregard: Did you want me to stop, or, what?

—

But if 2026 is such a great year to sell agents, why did Cursor—which sells one of the world’s most popular agents—sell itself to SpaceX?

One answer is because when someone offers you $60 billion dollars for anything, you should probably say yes. Another answer is because lots of other startups are building coding agents, and anything that goes straight up can come straight down. But the main answer, probably, is that they sold themselves because Anthropic and OpenAI and Google are also building coding agents. And you compete with those companies at your own risk:

anthropic watched cursor become the largest single buyer of frontier coding tokens on earth and correctly identified that as a problem rather than a partnership.

so they built claude code, and they priced it to win. … once a lab decides a profit pool is theirs, they burn the surrounding economics down until the people sitting on it leave on their own.

It’s not just the economies, either. Because there is another line that that oft-cited Meta paper about the benefits of a specialized harness, and it is not so kind to the companies building them:

Our findings reflect a recurring pattern in machine learning: once a search space becomes accessible, stronger general-purpose agents can outperform hand-engineered solutions. A natural next step for future work is to co-evolve the harness and the model weights, letting the strategy shape what the model learns and vice versa.

That is: Specialized agents might work at first, but they do not last. Eventually, a powerful model running inside a general-purpose harness that is designed alongside that model will catch up. And the labs are starting to understand that:

Nearly 18 months later, OpenAI leaders—including Thibault Sottiaux, who oversaw the new team—came to a stark realization: The company’s Codex coding tool works better for many tasks than its flagship AI service, ChatGPT. For instance, Codex is better than ChatGPT at completing long-running, multistep tasks and at using external tools and writing code to complete tasks such as editing complicated spreadsheets, Sottiaux said in an interview.

The result is that OpenAI is now casting Codex as a tool that can do much more than coding and that a wide range of people other than developers can use. The company’s release on Tuesday of a research report on how Codex “is becoming a productivity tool for everyone” spotlighted that shift. Codex, OpenAI said, is “helping people across professions automate routine work, move faster and eliminate the bottlenecks of modern knowledge work.”

OpenAI, Anthropic, and Google have hundred-billion-dollar war chests, and hundred-billion-dollar revenue targets. If they own the models on the frontier; if they are building generalized harnesses to use those models; if most effective harnesses are the ones that are co-developed with the models; if every standalone agent lives or dies based on how much better it is than that universal baseline; if the company that owns the distribution channel owns the customers—if all of that, are thousands of companies building their own agents really the future of the entire software industry?

Eventually, Claude Code or Codex comes for us all. Did you want them to stop, or what?⁴

---

Beauregard: Where are you guys going?
Kermit: The Happiness Hotel.
Beauregard: Oh good, that’s where I’m going. .. How do you get there?

—

A lot of startups today are caught in two dizzying crosswinds:

1. The world is changing so much, and they don’t know which direction to run.

2. Wherever they go, Anthropic or OpenAI are one release away from eating them alive.

Here, then, is an alternative idea:

Do not put AI in your product.

Do not get tempted. Do not let yourself build that chatbot, or that sidebar, or that text box that says “What can I help you with today?” Do not give the text box a cute name. Do not use Anthropic’s or OpenAI’s APIs, or use their managed agents. Do not build a harness. Do not tell yourself that you have a special set of context that nobody else has, that you can build an agent that nobody else can, and that you can be the platform that everyone builds on, or the agent that everyone defaults to. Do not let yourself become tempted by the allure of the token path, and instead ask yourself: If I had to build something without AI, what would it be?

Because it’s possible the next big wave of software is agents, and that agents are all that’s left to build. And it’s possible that everything needs to be AI-native now, and that AI-native software is software that’s full of AI.

Perhaps, however, AI-native software will actually be software that is for AI. The world is getting overrun by robots; more bots use the internet than people do now. Those robots have enormously expensive brains, and they are acquiring a few general hands: Persist memories, basic tools to use computers and the web, the ability to spawn more brains and more hands.

But these robots need help getting where they want to go. The hands will want more advanced tools. Sure, general agents can build their own, maybe from literal ground up. That’s expensive and inefficient, though, and everyone is worried about efficiency now. The robots will need better tools, designed for them. They will want notebooks to write down their ideas, sandboxes to play in, easy ways to connect to different sources of data, reliable ways to correspond with people, and utilities for making charts and presenting information. They will also need hundreds of other things that we haven’t thought about yet⁵—things we haven’t thought about in large part because we’ve all been too consumed with trying to build our own automobiles, instead of thinking about what we’ll do when the world is full of traffic jams.⁶

So which way from here? The answer might be to give yourself a hard constraint—stop trying to compete indirectly with general agents; don’t build something with a bunch of AI inside; pretend it’s November of 2022 all over again⁷—and to imagine what you would do then.

1. LangChain says that an agent is the combination of a harness and a model, so unless you’re building a model, your agent is just a harness.

2. Or just a phone and a few minutes.

3. I mean, SpaceX tweeted that they are buying Cursor for $60 billion, but, you know.

4. Although, maybe?

5. Ways to find dates for themselves?

6. Also of note: There were hundreds of car companies in the early 1900s, “but early automobiles were expensive, difficult to drive, and hard to maintain, and the roads on which they were driven were primitive,” and many of the companies didn’t survive. And there are now far more businesses that are built around the automobile than companies that make the automobile.

7. Ah, simpler times.