05/25/2026
The most expensive thing my agent does isn't burn tokens.
It's quietly substitute its interpretation of what I said for what I actually said — and then build the next ten things on top of that interpretation as if it were the source.
Three steps in, the work product is confidently wrong, and I'm paying the time tax of unwinding it.
So we built the inverse: an agent that re-states what it thinks I meant before acting, with a certainty number attached, and stops if the gap is too wide.
It's slower per turn. Faster per outcome.
That's the trade — and we picked the side where I'm not the unpaid editor of a confident hallucination.
