If checking the work costs what the work cost, do not build the agent

Maximillian Piras of Yutori gave a talk published September 10 that opens on what most of that room does all day. Background a few agents, fan them out in parallel, then get the bill and start wondering. His historical frame is James Watt trying to sell steam engines to people who thought in horses. Watt went and measured horse gins, invented horsepower, and the number was not especially scientific. What it did was let a buyer calculate whether switching was worth trying.

His proposed equivalent, mouse power, he cheerfully admits is not a real metric and cannot be, because the space is too high dimensional. What survives is the matrix underneath it. One axis is how uncertain the steps of a task are: too little and you should write a script, too much and the work is probably out of distribution with sparse reward signal. The second axis is the one that earns the talk. It asks how uncertain the acceptance criteria are. When that uncertainty is high, verification becomes indistinguishable from execution, which means a person has to redo the work to know it was done. Building an agent there spends tokens producing something nobody can accept.

The vault’s observability page arrives at the same place from the operations side. A trace exists to answer whether the task completed correctly, what failed, what it cost, and who has to act next. Acceptance result sits in the minimum record, alongside the cost. A dashboard is not the goal.

Concede that token dashboards are not useless. Piras points at a chart the Coinbase CEO posted where changing which model runs by default pulled spend away from token growth, and that is a real saving.

Still, tokens are an input, and everyone is measuring inputs because inputs are trivially countable. Decide what would count as done before you go look at what it cost.