Insights

Insights

Notes from systems we run ourselves, not a view from the outside.

  • Diagnosing a wrong answer: one minute instead of one week

    Citations, permission inheritance and confirmation gates are not safety decoration. They are what turns "the AI said something odd" from a week-long argument into a one-minute lookup.

  • Same question, 17 seconds to 7.9, without changing the model

    Latency is an orchestration property, not a model property. A routing change in our own system cut the same question by more than half, which means speed can be a requirement you specify rather than a number you accept.

  • Why twenty-plus agents in one system is cheaper

    An internal system we operate runs more than twenty agents plus scheduled jobs. Splitting them is what keeps instructions from diluting, makes permissions structural rather than prompted, and keeps a fix local to one configuration.

  • What an agent actually saves, calculated on your numbers

    Most time in a knowledge task goes on looking things up and moving results between systems. Both are measurable before anything is built — here is the arithmetic, with the assumptions left visible.