Writing

• Audit what's ambientTwelve hours into an overnight agent run, a test suite inherited production credentials and wiped my content tables. The gate I trusted never fired.
• The demand-generation taxSame company, same models, same reach - 36% daily engagement on one product, under 8% day-30 retention on the other. The difference is who generates the demand.
• The model-routing rubricModel choice is a P&L decision. The rubric I use to route every task to the cheapest model that clears the bar.
• Agents are the primary customerThe best new infra companies sell to software, not people. Chosen once by an agent, used indefinitely.
• The P&L lensStart from the number that has to move, then decide what AI is worth building. Not the other way around.