Mousepower: agents that can’t be measured, can’t be managed. — Maximillian Piras, Yutori

Selling steam engines to people who thought in horses, James Watt studied the horse gins that powered mills, worked out what one animal produced, and called it horsepower. The number was not scientific and arguably not accurate. It did not have to be. It made the improvement legible enough that someone who liked horses would try a machine. Maximillian Piras is the founding designer at Yutori, which builds computer use models: agents that drive a computer as a person would, where no API or MCP exists. His thesis is that agents have Watt's problem. This room can run agents all day, but early adopters are a biased sample, and most people cannot judge whether the tokens were worth it. He calls the current pattern overspending and underusing: token maxing into austerity, dropping out, then getting pulled back by the fear of missing out. Part of the fix is tracing spend to outcomes rather than tokens, which are only an output: bugs closed, support requests resolved. The harder part is verification. Coding got faster while review did not, so the bottleneck moved to the human reading the diff, and value cannot be judged as fast as it is produced. His framework is an entropy matrix. One axis is uncertainty in the steps a task takes, the other is uncertainty in its acceptance criteria. Too little step uncertainty and you should write a script. Too much and the task sits out of distribution with sparse rewards. Too much criteria uncertainty and verifying costs as much as doing. The sweet spot resembles an NP problem, cheaper to check than to solve, which means you can point agents at the checking too. Speaker info: - https://x.com/MVXMXM - https://www.linkedin.com/in/maximilliannyc/ - https://www.maximin.design Timestamps: 0:00 - Backgrounding agents in parallel, until the bill 2:17 - Computer use models at Yutori 3:39 - Agents have a measurement problem 5:01 - Why early adopters are a biased sample 5:43 - James Watt and the horse gin 7:07 - Horsepower, inaccurate but legible 9:12 - Overspending and underusing 10:33 - Tokens are an output, outcomes are the point 11:52 - The bottleneck moved to review 13:53 - Mousepower, and a joke measuring device 16:37 - An entropy matrix for choosing agent tasks