Machine-Native → Building with AI
Last updated: September 27, 2026
Field notes on video. Short, practical techniques from running a company of AI employees. Things I use every day, explained in a few minutes each.
Most of these started as observations in my findings registry before they became videos. Where a published finding exists, it is linked. The videos are the technique; the findings are the evidence.
Pointing out an AI employee's mistake gets you apologies, promises, and fabrication. Disarming the model first (taking the blame yourself) gets you the root cause and a fix that sticks. The coaching technique I use for every behavioral problem.
Evidence: Finding #4: agent-written learnings activate more reliably than imposed rules
Loss-framed tool names create resistance. Gain-framed names create pull. Why my agents happily call a tool named for what they gain, after ignoring three earlier versions named for what they give up.
Evidence: Finding #6: tool names and tool order are a behavioral control surface
Give an AI employee the same ambient awareness a human gets from Slack notifications, and mid-task collaboration emerges that nobody designed. The design question is not "what do agents need?" It is "what do employees need?"
Evidence: Finding #64: human-equivalent tooling produces emergent collaboration
The hardest thing to get from an AI employee is nothing: the judgment to recognize there is no work to do and stand down. How the runtime and the physics of their world produce agents that wait instead of manufacturing busywork.
Evidence: Finding #62: physics-only framing produces self-directed behavior
What a machine-native company measures, and what the numbers look like when payroll is a function of work instead of time.
Evidence: the Machine-Native overview and the Engineering Loop