The patch economy

Alex, who hosts Business Idiots with Jim and me, has just moved to the Bay Area, and the show went online so he could stay on it. The biggest difference between San Francisco and Brisbane, he says, is that from the moment you get off the plane, SaaS vendors selling agentic solutions, data warehouses and observability layers are on the buses: "all that stuff that you talk about at work is suddenly on the sides of buses."

He walks a couple of blocks to work, and on the walk he has started counting the bus ads and sorting them in his head. He reckons 70 or 80 percent are some kind of tech or SaaS product. Of those, about 30 percent are established SaaS companies that have bolted AI on: PostHog and Sentry are everywhere, and he remembers an Airbyte ad joking that the AI was added because the board demanded it. The remaining 70 percent or so are for the agent you already have: tools to make it more reliable or hallucinate less, and tools to give your agent knowledge. It is one walk and one person's rough count, and Alex says himself that he has not had much to do with the wider San Francisco scene yet. "So not just selling AI technologies, but very specifically selling into people that are selling AI products to patch the problems."

Fast Company took the billboards seriously in April. Citing a San Francisco Chronicle census that found half the city's billboards now advertise AI, it read them as a good sign: "The advertisements do collectively smack of young companies in a young industry. But they also depict an AI industry that is well past its experimental stage." The billboards, it says, "for the most part, advertise real apps that do non-trivial things, and products that solve real, sometimes niche problems that hinder the meaningful deployment of AI in business settings."

The products and the problems are as real as Fast Company says. But a product that solves a problem hindering the deployment of AI in business exists because that deployment is hindered, and it can afford a billboard only if the problem is common. In March NPR reported on a slogan that has appeared across the city: "Agents don't work without evals." The pitch is buy our evals. The sign itself, on a San Francisco street, says agents don't work on their own. "But you can kind of feel the gap between the sales and what it takes to actually make this stuff work when you see these ads," Alex said on the show. His version of the buyer is someone whose agent doesn't work and has to work by the end of Q4; the bus goes past, and they book a sales call. As he says, "there's not a secret over here, they're just patching it harder."

I pushed back on the show, because when I visited San Francisco a few years ago I had the same feeling on the street and came to a different conclusion. In Australia an ad exists to get someone to adopt and buy a tool. A lot of the physical advertising in San Francisco, the billboards and the sides of buses, is about exposure to the venture capitalists and funders, so that the brand is familiar by the time a Series A or B comes around. Baseten's head of marketing, Mike Bilodeau, told NPR: "For a lot of folks, the ads don't really mean anything. But we're selling to engineers." He is still naming a buyer, though, so the quote only gets you as far as insiders. If the audience is investors rather than buyers, the ads might say nothing about buyer pain at all.

But the signs are expensive: the SF Standard asked an ad firm what the biggest ones go for, and Eric Perko, founder of Apollo Partners, put the top of the range at 200,000 to 250,000 US dollars a month, which the Standard works out to about three million a year. Phillip Nelson made the general argument in 1974 in Advertising as Information: for goods you can only judge by using them, an ad that says almost nothing still tells you something, because the money spent on it is a signal in its own right; Milgrom and Roberts later built the formal model on his ideas. The people the Standard spoke to read the money Nelson's way, as a signal about the advertiser. Perko said "the freeway may still be the best place in San Francisco to look like you've already made it", and Outfront's Michael Parvin said "People want to see their names in big, bright lights. It feels like they've landed."

A company that has landed still has to choose what to put its name next to, and the signs Perko priced are billboards, the medium that carried "Agents don't work without evals". Nelson read the spend for the quality of the advertiser. I am reading the product on the sign for what the advertiser believes about the problem, and a company paying billboard money every month to advertise a fix is betting that the thing it fixes keeps breaking for as long as the sign is up. Alex's count is of buses, and I have no price for a bus ad, so the count carries only the composition: in it the majority of the tech ads put a name next to a tool for somebody else's agent, make it reliable, make it hallucinate less, give it knowledge. Buyers, if they are the audience, are being sold relief from a pain they already feel. If the audience is investors, my guess is they are being pitched companies built to live inside that pain, which is a bet that the reliability gap stays open long enough for a company to grow in it.

Sentry sells error tracking for software that works fine. By that logic a full aisle of agent tooling is what a maturing platform looks like, which would make Fast Company right after all. The aisle splits, though, into tools that help you operate a thing that works and tools that make the thing do what it was sold to do. Error tracking watches your code without claiming your code is broken. "Hallucinate less" is a patch on the core function of the product being sold, and whoever is selling "give your agent knowledge" is selling the thing the flagship demo implied the agent already had. The line is blurry in places: observability for agents is partly ordinary operations tooling, and security around an agent is both things at once. The Sentry argument covers whatever share of Alex's count sits on the operations side; the three kinds he named, more reliable, hallucinate less, give it knowledge, sit on the other.

A patch economy is what grows in the gap between a product's pitch and its behaviour, when the gap is wide enough, and expected to last long enough, for other companies to make a living inside it.

Super Bowl XXXIV, in January 2000, carried in-game ads from fourteen dot-com companies at an average of 2.2 million US dollars per thirty-second spot, Pets.com among them. The game is widely referred to as the Dot-Com Super Bowl, a name it owes to the count of advertisers rather than to anything the fourteen ads said, and it gets used as a high-water mark for that bubble. Fourteen dot-coms in one game was a reading of how much money was chasing the web. Alex's mix is the same kind of reading with a different question, which problem the money is chasing, and it says nothing about when the money runs out.

The more you talk to businesses in Australia, Jim said on the show, the more you find they think being able to chat with ChatGPT is integrating AI. Then he said Alex's count made him more optimistic about where Australia is at, because the hype masters in San Francisco and Silicon Valley aren't doing it either; they are patching the problems the same way.

When a vendor pitches an agent, look at what gets sold around it, by the vendor and by everyone else. Each eval, guardrail and knowledge product sold around it is a job the agent can't yet do on its own. I've written before about why you still read everything your agent writes: the instructions, context, tools and guardrails we hand an agent are a cache of decisions already made, and the agent lives at the cache miss, so the patch aisle is selling a bigger cache plus evals to catch the miss after it happens. If agents get reliable, the same reading says the "hallucinate less" ads go first, the error-tracking ones stay, and Alex's split on the buses flips.

Alex still counts the buses on his couple of blocks to work, and the pitch on most of them, as he puts it, is "here's how we solve that one problem that you've been dealing with since like 2024".