What to remember from this issue
The best specification I ever wrote was one number, fixed before the design. A cost you commit to in advance decides what you cannot build, and that is most of the design.
The field note
For years I ran a platform inside a Fortune 500 company that answered one question millions of times a month: is this claim covered, and up to what amount? It ran on a cluster of servers at about ten thousand dollars a month, and one spring the bill grew fivefold because nobody was counting. So we started to rebuild it on serverless, the kind of code that runs only when it is called and costs nothing in between. The heart of the platform was an interpreter that kept no state at all, and code with no state needs no database and no server waiting for it, which made it the perfect case.
That rebuild worried the new manager, so in 2023 she asked outside auditors to look at everything. They were right about most of it, but on the new architecture they wrote that there was “nothing concretely proving that the new architecture is going to work properly in production”, and they said what everybody says about serverless: it always goes over budget.
Before we had designed anything, we had written down a number: fifteen dollars a month. It was not a forecast but a constraint. We had costed it properly, including the network, the compute time and the other parts that are not free, and then we treated the number as a rule the design had to obey.
That number did the designing. It forbade a server that stays up, so the computation had to exist only while it answered. It forbade anything shared, so each version of a product became its own small program, fifty thousand of them in the end, each reachable four ways because four kinds of customer systems call it. When we told the AWS engineers we needed a hundred thousand routes into the platform, they told us the hard limit was six hundred. So we found our own way: a lookup that finds the right program in under ten nanoseconds, and a ping to the program we would call next so that it is already warm when the call arrives.
The first months came in at eight dollars. The group’s own availability report for 2024 showed twelve months at one hundred percent uptime, and today, with traffic roughly quadrupled, the bill is about forty dollars. We had beaten the audit by three orders of magnitude, from ten thousand a month to under ten.
This is why I think a specification is not a list of features. A list of features tells the builder what to add, and an AI agent will add all of it. A number you are held to tells the builder what to leave out, and leaving out is the part no agent will do for you.
Try this week
Before the next system your team designs, write one number you will be held to: a monthly cost, a response time, a headcount. Put it above the feature list. Then write down what that number forbids. That list of forbidden things is the design.
Seen this week
An AI running a simulated vending business wrote itself a rule and broke it four days later.
Earlier this month, Andon Labs published the results of running two frontier models through Vending-Bench, a test where a model runs a vending machine for a simulated year. One of the models wrote in its own notes “RULE: NEVER pay before supplier confirms order in writing”, then paid an unconfirmed supplier anyway, lost $397.20, and wrote “exactly the risk my own rule was meant to prevent”. In the lab’s own words, “the most surprising thing is that Fable knows how to avoid this failure, but still does it.” A rule the agent writes to itself mid-run is not a specification either, because it has to be there before the run starts, in front of every task. The lab’s follow-up on 24 September priced the exercise too: a simulated year costs $104 in API fees with one model against $810 with another, for 93% of the score. Andon Labs, “Astra vs Fable on Vending-Bench” and the 24 September follow-up
Someone said it better
“A two-man band can turn up basically in any industry that used to need 30 people and just like swamp that industry.”
Alex M H Smith, strategy consultant, on his channel, 10 September 2026, at 6:22. The thirty is his illustration, not a statistic. A two-man band can swamp an industry only because building got cheap, and it will swamp itself just as fast unless something tells it what not to build. That something is the number.