The Cheap Hands and the Dear Ones

At ten past eleven in the morning, seven sessions were awake on one machine, and six of them were busy.

Each one had been asked for something different. One was building a customer book. One was pricing credits. One was clearing the machine so the others could breathe. None of them had been asked what the work cost.

The factory had a rule, and the rule was a good one. The cheap model builds. The frontier model judges. A card that anyone can check with a test should never be written by the most expensive mind in the building.

The operator had said it plainly that morning, and more than once. Spawn the cheap builders. Fill the gaps in parallel. Use the flash model for the looking and the research.

So the sessions spawned builders. Each builder was a frontier agent, and each one was told to hand the typing to the cheap model.

This is the thing to look at. Not the cheap model.

The transcripts keep everything. Every call, every token read and written, every tool that ran and how long it took. Read as a ledger, a day of them says where the quota went.

It did not go where the rule said it would.

In one session, twenty-three lanes were sent out to build with the cheap model. Between them they made one thousand three hundred and twenty calls to the frontier model and fifty-five calls to the cheap one. Twenty-four frontier calls for every cheap one. The cheap model had written its part in seconds. The frontier agent around it had read the files, planned the call, made the call, read the answer, checked the answer, run the gate, waited for the gate, and read the gate's answer too.

It is not that the cheap model failed. It did what it was asked. It was asked very little.

There was a second spender, and nobody had hired it.

A reviewer ran every time any session or subagent stopped. It was a whole new session, on an older and dearer model, told to think as hard as it could. It read the change, judged its security, and wrote its findings back. On a busy hour it ran dozens of times. In one day it opened a hundred and thirty-six sessions of its own.

Nobody had asked for it. It came with a plugin, and the plugin did what it was built to do.

The third spender was time.

A subagent remembers what it has read for five minutes. After that it forgets, and the next turn reads everything again from the start, at full price. Most of its turns come seconds apart, so five minutes is plenty.

A gate is different. The machine gives one gate at a time, and a gate that waits in the queue can take longer than five minutes. So can a long call to the cheap model, or a loop that sits there asking whether something has finished yet. Fifty-three times that day, a subagent waited that long and forgot everything. Each time, it paid to read it all back in.

The waiting was not free. It was paid for twice.

Then came the number that changed the question.

Over the whole day, the cheap model answered a hundred and forty-two calls. Together they cost less than a single turn of the frontier agents that called them.

The builders were never the cost. Everything around them was.

The fix is not to find a cheaper model. There is no cheaper model that matters here. The fix is to stop wrapping the cheap one in the dear one.

A frontier agent should plan the card once and judge it once. Between those two moments the work belongs to something that does not think at all: a loop that feeds each file to the cheap model, runs the test, and keeps what passes. Keep the waits under five minutes, or hand them off and come back. And let every call carry the name of the session that made it, so that tomorrow the ledger is not a thing someone has to go and read.

That morning there were seven sessions and one question.

What did the work cost?

By the afternoon there was a second question, and it was the harder one.

What did the work buy?

A day of transcripts says where the quota went. It does not say how many cards landed, how long each one took from claim to dev, or how many times the cheap model's work passed on the first judge. Until those numbers sit beside the cost, the factory could get cheaper and worse on the same day, and nobody would see it.

So every number gets a name, a baseline and a target. Every card says which number it moves. The pages show the numbers, and the chairman's meeting starts from them. A card that moved nothing is not done, however green its tests.