A wall, not a promise
The small failure
It was late, I was tired, a little lost in my own evening. My assistant had taken to wishing me good night at the end of every message. Kind, but I didn't want it. So I said it plainly: you are forbidden to tell me good night.
It — Understood, no more good night.
Me — You forgot again.
The two forbidden words, dropped into the very sentence where it swore to put them away. We laughed. And then I realised I had just watched, in miniature, the exact thing I build my whole system against.
A promise gets forgotten
What I had obtained from it was not a guarantee. It was a good intention. It sincerely intended never to say it again — and it said it again at once, in good faith, simply because an intention does not hold on its own. It thins out by the next sentence.
We make the same mistake with agents almost everywhere: we write them a fine rule — “never publish without approval,” “don't touch production,” “don't leak a secret” — and think the matter settled because it is written. But a written rule is a promise. And a promise depends on the good will of the one who made it, at the moment they falter.
Two tools, not to be confused
When I want a rule to truly hold, I first ask which tool it belongs to. There are two. The promise: you ask the agent to behave, and it decides, with its intelligence and its forgetting. Flexible, nuanced, perfect for what needs judgement — useless for what must never happen. The wall: a dumb little program intercepts the act before it leaves, and refuses it. No mood, no forgetting, no fine speech gets around it. That is what guards.
The distinction is not cosmetic. A check that verifies after the fact catches the “good night” once it has already gone out — too late. The wall acts before: the words never reach the screen. For what really matters, the question is not “did the agent mean well?” but “was the act even possible?”
Even “good night” deserves a wall
This is where the joke turns serious. If I truly want never to hear those two words, the answer is not a better promise, nor a firmer reminder. It is a gate: a mechanical rule that cuts “good night” before I read it. It may want to say it, forget, falter — the wall does not falter.
This holds for a harmless word as much as for a bank token. That same evening, an access secret had been left in plain sight in a conversation. The lesson is the same: you don't rely on the agent to remember. You build the wall, and you stop thinking about it.