Maximillian Piras begins with the familiar habit of running many agents in parallel, then asks whether the resulting token bill creates enough value. He argues that token counts describe system output but do not tell customers which useful outcomes the agents produced.
Maximillian Piras uses James Watt's horsepower metric as an analogy for making unfamiliar technology legible through an existing mental model. For AI agents, he says spending should connect to outcomes such as resolved bugs or support requests, while verification costs must also be counted when code review becomes the bottleneck.
Maximillian Piras proposes evaluating potential agent tasks on two axes: uncertainty in the steps required and uncertainty in the acceptance criteria. The strongest candidates sit between rigid scripts and highly unpredictable work, with results that are easier to verify than to produce and can potentially be checked by another agent.
Watch on YouTube
